# --- Deep Learning: Modelos pre-entrenados ---
from keras.applications import vgg16, vgg19, ResNet50  # Arquitecturas CNN pre-entrenadas en ImageNet
from keras.models import Model                          # Clase para definir modelos personalizados
from keras.applications.imagenet_utils import preprocess_input  # Preprocesamiento estandar para CNNs

# --- Deep Learning: Utilidades de imagen ---
from tensorflow.keras.utils import load_img, img_to_array  # Carga y conversion de imagenes a arrays

# --- Procesamiento de imagenes ---
from PIL import Image  # Manipulacion basica de imagenes (Python Imaging Library)

# --- Computacion numerica y datos ---
import numpy as np   # Operaciones con arrays y matrices
import pandas as pd  # Manipulacion y analisis de datos tabulares

# --- Machine Learning: Metricas ---
from sklearn.metrics.pairwise import cosine_similarity  # Calculo de similitud coseno entre vectores

# --- Visualizacion ---
import matplotlib.pyplot as plt  # Generacion de graficos y visualizaciones

# --- Utilidades del sistema ---
import os  # Interaccion con el sistema de archivos

# --Extras--
import seaborn as sns

%%capture
!gdown 1iLOeNZw69iyYXa7QS5ZutN7ACUpkbL3x
!unzip images_fashion.zip
!mkdir images
!mv *.png images/

# Cargamos el modelo VGG19 pre-entrenado en ImageNet
vgg19_model = vgg19.VGG19(weights='imagenet')

# Quitar la capa de clasificacion (usamos fc2 como extractor de features)
feat_extractor = Model(inputs=vgg19_model.input, outputs=vgg19_model.get_layer("fc2").output)

# Vemos resumen de la arquitectura del modelo
print("Modelo: VGG19")
print("="*60)
feat_extractor.summary()

Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg19/vgg19_weights_tf_dim_ordering_tf_kernels.h5
574710816/574710816 ━━━━━━━━━━━━━━━━━━━━ 2s 0us/step
Modelo: VGG19
============================================================

Model: "functional"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ input_layer (InputLayer)        │ (None, 224, 224, 3)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_conv1 (Conv2D)           │ (None, 224, 224, 64)   │         1,792 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_conv2 (Conv2D)           │ (None, 224, 224, 64)   │        36,928 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_pool (MaxPooling2D)      │ (None, 112, 112, 64)   │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_conv1 (Conv2D)           │ (None, 112, 112, 128)  │        73,856 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_conv2 (Conv2D)           │ (None, 112, 112, 128)  │       147,584 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_pool (MaxPooling2D)      │ (None, 56, 56, 128)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv1 (Conv2D)           │ (None, 56, 56, 256)    │       295,168 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv2 (Conv2D)           │ (None, 56, 56, 256)    │       590,080 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv3 (Conv2D)           │ (None, 56, 56, 256)    │       590,080 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv4 (Conv2D)           │ (None, 56, 56, 256)    │       590,080 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_pool (MaxPooling2D)      │ (None, 28, 28, 256)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv1 (Conv2D)           │ (None, 28, 28, 512)    │     1,180,160 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv2 (Conv2D)           │ (None, 28, 28, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv3 (Conv2D)           │ (None, 28, 28, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv4 (Conv2D)           │ (None, 28, 28, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_pool (MaxPooling2D)      │ (None, 14, 14, 512)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv1 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv2 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv3 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv4 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_pool (MaxPooling2D)      │ (None, 7, 7, 512)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ flatten (Flatten)               │ (None, 25088)          │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ fc1 (Dense)                     │ (None, 4096)           │   102,764,544 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ fc2 (Dense)                     │ (None, 4096)           │    16,781,312 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 139,570,240 (532.42 MB)

 Trainable params: 139,570,240 (532.42 MB)

 Non-trainable params: 0 (0.00 B)

ls

images/  images_fashion.zip  __MACOSX/  sample_data/  style.csv  style.txt

imgs_path = "images/" # ruta

imgs_model_width, imgs_model_height = 224, 224 # tamaño de las imagenes 224x224 pixeles

nb_closest_images = 5 # cantidad de imagenes similares a recomendar

files = [imgs_path + x for x in os.listdir(imgs_path) if "png" in x]
print("total de imagenes:",len(files))

total de imagenes: 2184

# ============================================================================
# GALERIA DE 40 IMAGENES ALEATORIAS PARA SELECCION
# ============================================================================

import random

random.seed(42)  # Cambiar seed para ver diferentes imagenes
sample_indices = random.sample(range(len(files)), 40)

fig, axes = plt.subplots(5, 8, figsize=(20, 12), facecolor='#f8f9fa')

for i, ax in enumerate(axes.flat):
    idx = sample_indices[i]
    img = load_img(files[idx], target_size=(imgs_model_width, imgs_model_height))
    ax.imshow(img)
    ax.set_title(f'idx: {idx}', fontsize=9, fontweight='bold', color='#2c3e50')
    ax.axis('off')

plt.suptitle('GALERIA DE IMAGENES - Selecciona un indice para buscar similares',
             fontsize=16, fontweight='bold', color='#2c3e50', y=1.01)
plt.tight_layout()
plt.show()

print("\nIndices disponibles en esta muestra:")
print(sample_indices)
print(f"\nTotal de imagenes en el dataset: {len(files)}")

Indices disponibles en esta muestra:
[456, 102, 1126, 1003, 914, 571, 419, 356, 1728, 130, 122, 383, 895, 952, 2069, 108, 814, 1718, 902, 1839, 1139, 26, 653, 1731, 1393, 1138, 636, 881, 1378, 418, 379, 1556, 396, 1470, 1408, 1083, 177, 1881, 511, 1550]

Total de imagenes en el dataset: 2184

# Imagen seleccionada manualmente
idx = 1881  # Indice fijo seleccionado de la galeria
original = load_img(files[idx], target_size=(imgs_model_width, imgs_model_height))
plt.imshow(original)
plt.show()
print(f"Indice: {idx}")
print("Imagen cargada exitosamente!")

Indice: 1881
Imagen cargada exitosamente!

# convertir PIL image a numpy array
numpy_image = img_to_array(original)

# convertir imagen a batch de imagenes para entrenamiento mas eficiente
image_batch = np.expand_dims(numpy_image, axis=0)

# preparamos la imagen para la VGG16
processed_image = preprocess_input(image_batch.copy())

print(f"Imagen procesada: idx = {idx}")
print(f"Archivo: {files[idx]}")
print(f"Image batch size: {image_batch.shape}")

processed_image

Imagen procesada: idx = 1881
Archivo: images/3_2_045.png
Image batch size: (1, 224, 224, 3)

array([[[[126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         ...,
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ]],

        [[126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         ...,
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ]],

        [[126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         ...,
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ],
         [126.061, 113.221, 106.32 ]],

        ...,

        [[127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         ...,
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ]],

        [[127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         ...,
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ]],

        [[127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         ...,
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ],
         [127.061, 114.221, 107.32 ]]]], dtype=float32)

# ============================================================================
# VISUALIZACIÓN DETALLADA DE LA TRANSFORMACIÓN
# ============================================================================

fig = plt.figure(figsize=(18, 12))

# 1. Imagen Original
ax1 = fig.add_subplot(3, 4, 1)
ax1.imshow(original)
ax1.set_title('Imagen Original', fontsize=12, fontweight='bold')
ax1.axis('off')

# 2-4. Canales RGB separados
channel_names = ['Canal Rojo (R)', 'Canal Verde (G)', 'Canal Azul (B)']
cmaps = ['Reds', 'Greens', 'Blues']

for i in range(3):
    ax = fig.add_subplot(3, 4, i+2)
    im = ax.imshow(numpy_image[:,:,i], cmap=cmaps[i])
    ax.set_title(channel_names[i], fontsize=11, fontweight='bold')
    ax.axis('off')
    plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)

# 5. Histograma de valores de píxeles (antes del preprocesamiento)
ax5 = fig.add_subplot(3, 4, 5)
colors = ['red', 'green', 'blue']
for i, color in enumerate(colors):
    ax5.hist(numpy_image[:,:,i].flatten(), bins=50, alpha=0.5, color=color, label=f'Canal {color[0].upper()}')
ax5.set_title('Histograma RGB (Original)', fontsize=12, fontweight='bold')
ax5.set_xlabel('Valor de Pixel (0-255)')
ax5.set_ylabel('Frecuencia')
ax5.legend()
ax5.grid(True, alpha=0.3)

# 6. Imagen preprocesada (normalizada para visualización)
ax6 = fig.add_subplot(3, 4, 6)
# Normalizar para visualización (el preprocesamiento puede dar valores negativos)
processed_vis = processed_image[0] - processed_image[0].min()
processed_vis = (processed_vis / processed_vis.max() * 255).astype(np.uint8)
ax6.imshow(processed_vis)
ax6.set_title('Imagen Preprocesada\n(Normalizada para visualizacion)', fontsize=11, fontweight='bold')
ax6.axis('off')

# 7. Histograma después del preprocesamiento
ax7 = fig.add_subplot(3, 4, 7)
for i, color in enumerate(colors):
    ax7.hist(processed_image[0,:,:,i].flatten(), bins=50, alpha=0.5, color=color, label=f'Canal {color[0].upper()}')
ax7.set_title('Histograma RGB (Preprocesado)', fontsize=12, fontweight='bold')
ax7.set_xlabel('Valor de Pixel (centrado)')
ax7.set_ylabel('Frecuencia')
ax7.legend()
ax7.grid(True, alpha=0.3)

# 8. Heatmap de intensidad promedio
ax8 = fig.add_subplot(3, 4, 8)
intensity = np.mean(numpy_image, axis=2)
im8 = ax8.imshow(intensity, cmap='hot')
ax8.set_title('Mapa de Intensidad', fontsize=12, fontweight='bold')
ax8.axis('off')
plt.colorbar(im8, ax=ax8, fraction=0.046, pad=0.04)

# 9-11. Vista de matriz numérica (esquina superior izquierda 8x8)
for i in range(3):
    ax = fig.add_subplot(3, 4, 9+i)
    patch = numpy_image[:8, :8, i]
    im = ax.imshow(patch, cmap=cmaps[i], vmin=0, vmax=255)
    ax.set_title(f'Matriz {["R","G","B"][i]} (8x8 pixeles)', fontsize=10, fontweight='bold')
    # Añadir valores numéricos
    for y in range(8):
        for x in range(8):
            ax.text(x, y, f'{int(patch[y,x])}', ha='center', va='center',
                    fontsize=6, color='white' if patch[y,x] < 128 else 'black')
    ax.set_xticks([])
    ax.set_yticks([])

# 12. Estadísticas
ax12 = fig.add_subplot(3, 4, 12)
ax12.axis('off')
stats_text = f"""
ESTADISTICAS DE LA IMAGEN

Shape Original: {numpy_image.shape}
Shape Batch: {image_batch.shape}
Shape Procesada: {processed_image.shape}

VALORES ORIGINALES (0-255):
   R: mean={numpy_image[:,:,0].mean():.1f}, std={numpy_image[:,:,0].std():.1f}
   G: mean={numpy_image[:,:,1].mean():.1f}, std={numpy_image[:,:,1].std():.1f}
   B: mean={numpy_image[:,:,2].mean():.1f}, std={numpy_image[:,:,2].std():.1f}

VALORES PREPROCESADOS:
   R: mean={processed_image[0,:,:,0].mean():.1f}, std={processed_image[0,:,:,0].std():.1f}
   G: mean={processed_image[0,:,:,1].mean():.1f}, std={processed_image[0,:,:,1].std():.1f}
   B: mean={processed_image[0,:,:,2].mean():.1f}, std={processed_image[0,:,:,2].std():.1f}

Total pixeles: {numpy_image.shape[0] * numpy_image.shape[1]:,}
"""
ax12.text(0.1, 0.95, stats_text, transform=ax12.transAxes, fontsize=10,
          verticalalignment='top', fontfamily='monospace',
          bbox=dict(boxstyle='round', facecolor='#f0f0f0', alpha=0.8))

plt.suptitle('ANALISIS DETALLADO DE TRANSFORMACION DE IMAGEN PARA CNN',
             fontsize=16, fontweight='bold', y=1.02)
plt.tight_layout()
plt.show()

print("\nImagen convertida y preprocesada exitosamente para la CNN!")

Imagen convertida y preprocesada exitosamente para la CNN!

# obtenemos los features (embeddings) de las imagenes pasandolas por la VGG19
img_features = feat_extractor.predict(processed_image)

print("features successfully extracted!")
print("number of image features:",img_features.size)
img_features

1/1 ━━━━━━━━━━━━━━━━━━━━ 3s 3s/step
features successfully extracted!
number of image features: 4096

array([[0.       , 0.       , 1.282129 , ..., 0.       , 0.       ,
        3.2349215]], dtype=float32)

img_features.shape

(1, 4096)

# repetimos el mismo proceso para todas las imagenes y guardamos los batch en una lista para entregarselos procesados a la VGG19
importedImages = []

for f in files:
    filename = f
    original = load_img(filename, target_size=(224, 224))
    numpy_image = img_to_array(original)
    image_batch = np.expand_dims(numpy_image, axis=0)

    importedImages.append(image_batch)

images = np.vstack(importedImages)

processed_imgs = preprocess_input(images.copy())

# obtenemos los features para cada imagen con la CNN
imgs_features = feat_extractor.predict(processed_imgs)

print("features extraidos exitosamente!")
imgs_features.shape

69/69 ━━━━━━━━━━━━━━━━━━━━ 28s 233ms/step
features extraidos exitosamente!

(2184, 4096)

# computa similaridad coseno entre los features de las imagenes
cosSimilarities = cosine_similarity(imgs_features)

# guardamos los resultados en un dataframe
cos_similarities_df = pd.DataFrame(cosSimilarities, columns=files, index=files)
cos_similarities_df #.head()

# ============================================================================
# VISUALIZACION DE LA MATRIZ DE SIMILITUD COSENO
# ============================================================================

print(f"Imagen de referencia: idx = {idx}")
print(f"Archivo: {files[idx]}")
print("="*60 + "\n")

fig = plt.figure(figsize=(18, 14))

# 1. Heatmap de la matriz de similitud completa
ax1 = fig.add_subplot(2, 2, 1)
# Tomamos una muestra para que sea visible (cada N imagenes)
sample_step = max(1, len(files) // 50)
sample_indices = list(range(0, len(files), sample_step))
similarity_sample = cosSimilarities[np.ix_(sample_indices, sample_indices)]

im1 = ax1.imshow(similarity_sample, cmap='viridis', aspect='auto')
ax1.set_title('Heatmap de Similitud Coseno\n(Muestra de imagenes)', fontsize=14, fontweight='bold')
ax1.set_xlabel('Indice de Imagen', fontsize=11)
ax1.set_ylabel('Indice de Imagen', fontsize=11)
cbar1 = plt.colorbar(im1, ax=ax1, fraction=0.046, pad=0.04)
cbar1.set_label('Similitud Coseno', fontsize=10)

# 2. Distribucion de similitudes (histograma)
ax2 = fig.add_subplot(2, 2, 2)
# Extraer triangulo superior (sin diagonal) para evitar duplicados
upper_triangle = cosSimilarities[np.triu_indices_from(cosSimilarities, k=1)]
ax2.hist(upper_triangle, bins=80, color='#3498db', alpha=0.7, edgecolor='black', linewidth=0.5)
ax2.axvline(upper_triangle.mean(), color='red', linestyle='--', linewidth=2, label=f'Media: {upper_triangle.mean():.4f}')
ax2.axvline(np.median(upper_triangle), color='orange', linestyle='--', linewidth=2, label=f'Mediana: {np.median(upper_triangle):.4f}')
ax2.set_title('Distribucion de Similitudes entre Imagenes', fontsize=14, fontweight='bold')
ax2.set_xlabel('Score de Similitud Coseno', fontsize=11)
ax2.set_ylabel('Frecuencia', fontsize=11)
ax2.legend(fontsize=10)
ax2.grid(True, alpha=0.3)

# 3. Top 10 pares mas similares (excluyendo la misma imagen)
ax3 = fig.add_subplot(2, 2, 3)
# Encontrar los pares mas similares
n_images = len(files)
pairs = []
for i in range(n_images):
    for j in range(i+1, n_images):
        pairs.append((i, j, cosSimilarities[i, j]))

# Ordenar por similitud descendente
pairs_sorted = sorted(pairs, key=lambda x: x[2], reverse=True)[:15]

# Graficar barras horizontales
pair_labels = [f'Img {p[0]} - Img {p[1]}' for p in pairs_sorted]
pair_scores = [p[2] for p in pairs_sorted]
colors_gradient = plt.cm.RdYlGn(np.linspace(0.9, 0.5, len(pair_scores)))

bars = ax3.barh(range(len(pair_labels)), pair_scores, color=colors_gradient, edgecolor='black', linewidth=0.5)
ax3.set_yticks(range(len(pair_labels)))
ax3.set_yticklabels(pair_labels, fontsize=9)
ax3.set_xlabel('Similitud Coseno', fontsize=11)
ax3.set_title('Top 15 Pares de Imagenes Mas Similares', fontsize=14, fontweight='bold')
ax3.invert_yaxis()
ax3.grid(True, axis='x', alpha=0.3)

# Añadir valores en las barras
for bar, score in zip(bars, pair_scores):
    ax3.text(bar.get_width() + 0.005, bar.get_y() + bar.get_height()/2,
             f'{score:.4f}', va='center', fontsize=8, fontweight='bold')

# 4. Estadisticas y boxplot
ax4 = fig.add_subplot(2, 2, 4)

# Boxplot con violin plot combinado
parts = ax4.violinplot([upper_triangle], positions=[1], showmeans=True, showmedians=True)
parts['bodies'][0].set_facecolor('#3498db')
parts['bodies'][0].set_alpha(0.7)

# Boxplot superpuesto
bp = ax4.boxplot([upper_triangle], positions=[1], widths=0.15, patch_artist=True)
bp['boxes'][0].set_facecolor('#2ecc71')
bp['boxes'][0].set_alpha(0.7)

ax4.set_ylabel('Similitud Coseno', fontsize=11)
ax4.set_title('Distribucion de Similitudes\n(Violin + Box Plot)', fontsize=14, fontweight='bold')
ax4.set_xticks([1])
ax4.set_xticklabels(['Todas las\ncomparaciones'], fontsize=11)
ax4.grid(True, axis='y', alpha=0.3)

# Añadir estadisticas como texto
stats_text = f"""
ESTADISTICAS GLOBALES
---------------------
Total imagenes: {n_images:,}
Total comparaciones: {len(upper_triangle):,}

Similitud minima: {upper_triangle.min():.4f}
Similitud maxima: {upper_triangle.max():.4f}
Media: {upper_triangle.mean():.4f}
Mediana: {np.median(upper_triangle):.4f}
Desv. Estandar: {upper_triangle.std():.4f}

Percentil 25: {np.percentile(upper_triangle, 25):.4f}
Percentil 75: {np.percentile(upper_triangle, 75):.4f}
Percentil 95: {np.percentile(upper_triangle, 95):.4f}
"""
ax4.text(1.6, 0.5, stats_text, transform=ax4.get_xaxis_transform(), fontsize=9,
         verticalalignment='center', fontfamily='monospace',
         bbox=dict(boxstyle='round', facecolor='#f8f9fa', alpha=0.9, edgecolor='gray'))

ax4.set_xlim(0.5, 2.8)

plt.suptitle('ANALISIS DE SIMILITUD COSENO ENTRE IMAGENES',
             fontsize=18, fontweight='bold', y=1.02)
plt.tight_layout()
plt.show()

print(f"\nMatriz de similitud calculada exitosamente!")
print(f"Dimensiones: {cosSimilarities.shape[0]} x {cosSimilarities.shape[1]} = {cosSimilarities.shape[0]**2:,} comparaciones")

Imagen de referencia: idx = 1881
Archivo: images/3_2_045.png
============================================================

Matriz de similitud calculada exitosamente!
Dimensiones: 2184 x 2184 = 4,769,856 comparaciones

# esta funcion recupera las imagenes más similares dada una imagen entregada por el usuario
def retrieve_most_similar_products(given_img):

    print("-----------------------------------------------------------------------")
    print("producto escogido:")

    original = load_img(given_img, target_size=(imgs_model_width, imgs_model_height))
    plt.imshow(original)
    plt.show()

    print("-----------------------------------------------------------------------")
    print("productos más similares:")

    closest_imgs = cos_similarities_df[given_img].sort_values(ascending=False)[1:nb_closest_images+1].index
    closest_imgs_scores = cos_similarities_df[given_img].sort_values(ascending=False)[1:nb_closest_images+1]

    for i in range(0,len(closest_imgs)):
        original = load_img(closest_imgs[i], target_size=(imgs_model_width, imgs_model_height))
        plt.imshow(original)
        plt.show()
        print("score de similaridad : ",closest_imgs_scores[i])

# ============================================================================
# VISUALIZACION DE RECOMENDACIONES CON ANALISIS DE FEATURES
# ============================================================================

query_image = files[idx]

print(f"Indice seleccionado: {idx}")
print(f"Archivo: {query_image}")

# Obtener las imagenes mas similares y sus scores
closest_imgs = cos_similarities_df[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
closest_imgs_paths = closest_imgs.index.tolist()
closest_imgs_scores = closest_imgs.values

# Obtener indices para extraer features
query_idx = files.index(query_image)
rec_indices = [files.index(p) for p in closest_imgs_paths]

# Colores para ranking
ranking_colors = ['#27ae60', '#2ecc71', '#f1c40f', '#e67e22', '#e74c3c']

# ============================================================================
# FIGURA 1: Imagen de Consulta
# ============================================================================
fig1, ax1 = plt.subplots(figsize=(6, 6), facecolor='#f8f9fa')
query_img = load_img(query_image, target_size=(imgs_model_width, imgs_model_height))
ax1.imshow(query_img)
ax1.set_title('IMAGEN DE CONSULTA', fontsize=18, fontweight='bold', color='#2c3e50', pad=20)
ax1.axis('off')

# Borde elegante
for spine in ax1.spines.values():
    spine.set_visible(True)
    spine.set_color('#3498db')
    spine.set_linewidth(5)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 2: Top 5 Recomendaciones
# ============================================================================
print("\n" + "="*70)
print("TOP 5 PRODUCTOS RECOMENDADOS")
print("="*70 + "\n")

fig2, axes2 = plt.subplots(1, nb_closest_images, figsize=(18, 4), facecolor='#f8f9fa')

for i in range(nb_closest_images):
    ax = axes2[i]
    rec_img = load_img(closest_imgs_paths[i], target_size=(imgs_model_width, imgs_model_height))
    ax.imshow(rec_img)
    ax.axis('off')

    score = closest_imgs_scores[i]
    ax.set_title(f'#{i+1} | Similitud: {score:.4f}', fontsize=13, fontweight='bold',
                 color=ranking_colors[i], pad=15)

    # Borde con color de ranking
    for spine in ax.spines.values():
        spine.set_visible(True)
        spine.set_color(ranking_colors[i])
        spine.set_linewidth(4)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 3: Comparacion de Features (Barras)
# ============================================================================
print("\n" + "="*70)
print("ANALISIS COMPARATIVO DE FEATURES")
print("="*70 + "\n")

fig3, ax3 = plt.subplots(figsize=(12, 5), facecolor='#f8f9fa')

bars = ax3.barh(range(nb_closest_images), closest_imgs_scores,
                color=ranking_colors, edgecolor='#2c3e50', height=0.6, linewidth=1.5)

ax3.set_yticks(range(nb_closest_images))
ax3.set_yticklabels([f'Recomendacion #{i+1}' for i in range(nb_closest_images)], fontsize=12, fontweight='bold')
ax3.set_xlabel('Score de Similitud Coseno', fontsize=13, fontweight='bold')
ax3.set_title('RANKING DE SIMILITUD', fontsize=16, fontweight='bold', color='#2c3e50', pad=20)
ax3.set_xlim(0, 1)
ax3.invert_yaxis()
ax3.set_facecolor('#fafafa')
ax3.grid(True, axis='x', alpha=0.4, linestyle='--')
ax3.spines['top'].set_visible(False)
ax3.spines['right'].set_visible(False)

for bar, score in zip(bars, closest_imgs_scores):
    ax3.text(score + 0.02, bar.get_y() + bar.get_height()/2,
             f'{score:.4f}', va='center', fontsize=12, fontweight='bold')

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 4: Comparacion de Vectores de Features
# ============================================================================
fig4, axes4 = plt.subplots(2, 3, figsize=(16, 10), facecolor='#f8f9fa')

# Features de la imagen query
query_features = imgs_features[query_idx]

# Plot 1: Distribucion de features de la imagen query
ax = axes4[0, 0]
ax.hist(query_features, bins=50, color='#3498db', alpha=0.7, edgecolor='black', linewidth=0.5)
ax.set_title('Distribucion de Features\n(Imagen Query)', fontsize=12, fontweight='bold')
ax.set_xlabel('Valor del Feature')
ax.set_ylabel('Frecuencia')
ax.axvline(query_features.mean(), color='red', linestyle='--', linewidth=2, label=f'Media: {query_features.mean():.2f}')
ax.legend(fontsize=9)
ax.grid(True, alpha=0.3)

# Plot 2: Primeros 100 features comparados (Query vs Top 1)
print("\n")
ax = axes4[0, 1]
feature_range = range(100)
ax.plot(feature_range, query_features[:100], 'b-', alpha=0.7, linewidth=1.5, label='Query')
ax.plot(feature_range, imgs_features[rec_indices[0]][:100], 'g--', alpha=0.7, linewidth=1.5, label='Top #1')
ax.set_title('Comparacion de Features (0-100)\nQuery vs Top #1', fontsize=12, fontweight='bold')
ax.set_xlabel('Indice del Feature')
ax.set_ylabel('Valor')
ax.legend(fontsize=10)
ax.grid(True, alpha=0.3)

# Plot 3: Heatmap de correlacion entre query y recomendaciones
print("\n")
ax = axes4[0, 2]
correlation_matrix = np.zeros((6, 6))
all_features = [query_features] + [imgs_features[idx] for idx in rec_indices]
labels = ['Query'] + [f'#{i+1}' for i in range(nb_closest_images)]

for i in range(6):
    for j in range(6):
        correlation_matrix[i, j] = cosine_similarity([all_features[i]], [all_features[j]])[0][0]

im = ax.imshow(correlation_matrix, cmap='RdYlGn', vmin=0.5, vmax=1)
ax.set_xticks(range(6))
ax.set_yticks(range(6))
ax.set_xticklabels(labels, fontsize=10, fontweight='bold')
ax.set_yticklabels(labels, fontsize=10, fontweight='bold')
ax.set_title('Matriz de Similitud\nQuery vs Recomendaciones', fontsize=12, fontweight='bold')
plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)

# Añadir valores en el heatmap
for i in range(6):
    for j in range(6):
        ax.text(j, i, f'{correlation_matrix[i,j]:.2f}', ha='center', va='center',
                fontsize=9, fontweight='bold', color='white' if correlation_matrix[i,j] > 0.75 else 'black')

# Plot 4: Activaciones mas altas (Top features)
ax = axes4[1, 0]
top_k = 20
top_indices = np.argsort(query_features)[-top_k:][::-1]
top_values = query_features[top_indices]

colors_bar = plt.cm.Blues(np.linspace(0.4, 0.9, top_k))
ax.barh(range(top_k), top_values, color=colors_bar, edgecolor='black', linewidth=0.5)
ax.set_yticks(range(top_k))
ax.set_yticklabels([f'F-{idx}' for idx in top_indices], fontsize=8)
ax.set_xlabel('Valor de Activacion')
ax.set_title(f'Top {top_k} Features con Mayor Activacion\n(Imagen Query)', fontsize=12, fontweight='bold')
ax.invert_yaxis()
ax.grid(True, axis='x', alpha=0.3)

# Plot 5: Diferencia de features entre Query y cada recomendacion
ax = axes4[1, 1]
differences = []
for i, rec_idx in enumerate(rec_indices):
    diff = np.mean(np.abs(query_features - imgs_features[rec_idx]))
    differences.append(diff)

bars = ax.bar(range(nb_closest_images), differences, color=ranking_colors, edgecolor='black', linewidth=1.5)
ax.set_xticks(range(nb_closest_images))
ax.set_xticklabels([f'#{i+1}' for i in range(nb_closest_images)], fontsize=11, fontweight='bold')
ax.set_ylabel('Diferencia Promedio Absoluta')
ax.set_title('Diferencia de Features\nQuery vs Recomendaciones', fontsize=12, fontweight='bold')
ax.grid(True, axis='y', alpha=0.3)

# Corregido: posicionar texto justo encima de cada barra
for i, (bar, diff) in enumerate(zip(bars, differences)):
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{diff:.2f}', ha='center', va='bottom', fontsize=10, fontweight='bold')

# Plot 6: Estadisticas resumen
ax = axes4[1, 2]
ax.axis('off')

stats_text = f"""
ESTADISTICAS DE FEATURES
{'='*40}

IMAGEN QUERY:
  - Dimension del vector: {len(query_features):,}
  - Media: {query_features.mean():.4f}
  - Std: {query_features.std():.4f}
  - Min: {query_features.min():.4f}
  - Max: {query_features.max():.4f}
  - Features activos (>0): {np.sum(query_features > 0):,}

SIMILITUDES:
  - Mejor match: #{1} ({closest_imgs_scores[0]:.4f})
  - Peor match: #{nb_closest_images} ({closest_imgs_scores[-1]:.4f})
  - Rango: {closest_imgs_scores[0] - closest_imgs_scores[-1]:.4f}

DIFERENCIAS PROMEDIO:
  - Menor diferencia: #{np.argmin(differences)+1} ({min(differences):.2f})
  - Mayor diferencia: #{np.argmax(differences)+1} ({max(differences):.2f})
"""

ax.text(0.1, 0.95, stats_text, transform=ax.transAxes, fontsize=11,
        verticalalignment='top', fontfamily='monospace',
        bbox=dict(boxstyle='round', facecolor='#e8f4f8', alpha=0.9, edgecolor='#3498db', linewidth=2))

plt.suptitle('ANALISIS DETALLADO DE FEATURES CNN', fontsize=18, fontweight='bold', color='#2c3e50', y=1.02)
plt.tight_layout()
plt.show()

# ============================================================================
# RESUMEN TABULAR
# ============================================================================
print("\n")
print("+" + "="*62 + "+")
print("|{:^62}|".format("RESUMEN DE RECOMENDACIONES"))
print("+" + "="*62 + "+")
print("| {:<10} | {:<15} | {:<30} |".format("Ranking", "Similitud", "Archivo"))
print("+" + "-"*62 + "+")
for i, (path, score) in enumerate(zip(closest_imgs_paths, closest_imgs_scores)):
    filename = path.split('/')[-1]
    print("| {:<10} | {:<15.4f} | {:<30} |".format(f"#{i+1}", score, filename))
print("+" + "="*62 + "+")

Indice seleccionado: 1881
Archivo: images/3_2_045.png

======================================================================
TOP 5 PRODUCTOS RECOMENDADOS
======================================================================

======================================================================
ANALISIS COMPARATIVO DE FEATURES
======================================================================


+==============================================================+
|                  RESUMEN DE RECOMENDACIONES                  |
+==============================================================+
| Ranking    | Similitud       | Archivo                        |
+--------------------------------------------------------------+
| #1         | 0.6318          | 3_2_066.png                    |
| #2         | 0.6237          | 4_2_030.png                    |
| #3         | 0.5985          | 4_2_029.png                    |
| #4         | 0.5978          | 0_2_006.png                    |
| #5         | 0.5943          | 5_2_059.png                    |
+==============================================================+

# Cargamos el modelo VGG16 pre-entrenado en ImageNet
vgg16_model = vgg16.VGG16(weights='imagenet')

# Quitar la capa de clasificacion (usamos fc2 como extractor de features)
feat_extractor_vgg16 = Model(inputs=vgg16_model.input, outputs=vgg16_model.get_layer("fc2").output)

# Vemos resumen de la arquitectura del modelo
print("Modelo: VGG16")
print("="*60)
feat_extractor_vgg16.summary()

Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels.h5
553467096/553467096 ━━━━━━━━━━━━━━━━━━━━ 2s 0us/step
Modelo: VGG16
============================================================

Model: "functional_1"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ input_layer_1 (InputLayer)      │ (None, 224, 224, 3)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_conv1 (Conv2D)           │ (None, 224, 224, 64)   │         1,792 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_conv2 (Conv2D)           │ (None, 224, 224, 64)   │        36,928 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block1_pool (MaxPooling2D)      │ (None, 112, 112, 64)   │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_conv1 (Conv2D)           │ (None, 112, 112, 128)  │        73,856 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_conv2 (Conv2D)           │ (None, 112, 112, 128)  │       147,584 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block2_pool (MaxPooling2D)      │ (None, 56, 56, 128)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv1 (Conv2D)           │ (None, 56, 56, 256)    │       295,168 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv2 (Conv2D)           │ (None, 56, 56, 256)    │       590,080 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_conv3 (Conv2D)           │ (None, 56, 56, 256)    │       590,080 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block3_pool (MaxPooling2D)      │ (None, 28, 28, 256)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv1 (Conv2D)           │ (None, 28, 28, 512)    │     1,180,160 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv2 (Conv2D)           │ (None, 28, 28, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_conv3 (Conv2D)           │ (None, 28, 28, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block4_pool (MaxPooling2D)      │ (None, 14, 14, 512)    │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv1 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv2 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_conv3 (Conv2D)           │ (None, 14, 14, 512)    │     2,359,808 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ block5_pool (MaxPooling2D)      │ (None, 7, 7, 512)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ flatten (Flatten)               │ (None, 25088)          │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ fc1 (Dense)                     │ (None, 4096)           │   102,764,544 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ fc2 (Dense)                     │ (None, 4096)           │    16,781,312 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 134,260,544 (512.16 MB)

 Trainable params: 134,260,544 (512.16 MB)

 Non-trainable params: 0 (0.00 B)

# obtenemos los features (embeddings) de las imagenes pasandolas por la VGG16
img_features_vgg16 = feat_extractor_vgg16.predict(processed_image)

print("Modelo: VGG16")
print("Features successfully extracted!")
print("Number of image features:", img_features_vgg16.size)
img_features_vgg16

1/1 ━━━━━━━━━━━━━━━━━━━━ 1s 609ms/step
Modelo: VGG16
Features successfully extracted!
Number of image features: 4096

array([[0.        , 0.06179556, 0.        , ..., 0.        , 0.        ,
        0.38277382]], dtype=float32)

img_features_vgg16.shape

(1, 4096)

# repetimos el mismo proceso para todas las imagenes y guardamos los batch en una lista para entregarselos procesados a la VGG16
importedImages = []

for f in files:
    filename = f
    original = load_img(filename, target_size=(224, 224))
    numpy_image = img_to_array(original)
    image_batch = np.expand_dims(numpy_image, axis=0)
    importedImages.append(image_batch)

images = np.vstack(importedImages)
processed_imgs = preprocess_input(images.copy())

# obtenemos los features para cada imagen con la CNN VGG16
imgs_features_vgg16 = feat_extractor_vgg16.predict(processed_imgs)

print("Modelo: VGG16")
print("Features extraidos exitosamente!")
imgs_features_vgg16.shape

69/69 ━━━━━━━━━━━━━━━━━━━━ 11s 152ms/step
Modelo: VGG16
Features extraidos exitosamente!

(2184, 4096)

# computa similaridad coseno entre los features de las imagenes (VGG16)
cosSimilarities_vgg16 = cosine_similarity(imgs_features_vgg16)

# guardamos los resultados en un dataframe
cos_similarities_df_vgg16 = pd.DataFrame(cosSimilarities_vgg16, columns=files, index=files)

print("Modelo: VGG16")
print(f"Matriz de similitud: {cosSimilarities_vgg16.shape}")
cos_similarities_df_vgg16

Modelo: VGG16
Matriz de similitud: (2184, 2184)

# ============================================================================
# VISUALIZACION DE LA MATRIZ DE SIMILITUD COSENO - VGG16
# ============================================================================

print("Modelo: VGG16")
print(f"Imagen de referencia: idx = {idx}")
print(f"Archivo: {files[idx]}")
print("="*60 + "\n")

fig = plt.figure(figsize=(18, 14))

# 1. Heatmap de la matriz de similitud completa
ax1 = fig.add_subplot(2, 2, 1)
# Tomamos una muestra para que sea visible (cada N imagenes)
sample_step = max(1, len(files) // 50)
sample_indices = list(range(0, len(files), sample_step))
similarity_sample = cosSimilarities_vgg16[np.ix_(sample_indices, sample_indices)]

im1 = ax1.imshow(similarity_sample, cmap='viridis', aspect='auto')
ax1.set_title('Heatmap de Similitud Coseno\n(Muestra de imagenes)', fontsize=14, fontweight='bold')
ax1.set_xlabel('Indice de Imagen', fontsize=11)
ax1.set_ylabel('Indice de Imagen', fontsize=11)
cbar1 = plt.colorbar(im1, ax=ax1, fraction=0.046, pad=0.04)
cbar1.set_label('Similitud Coseno', fontsize=10)

# 2. Distribucion de similitudes (histograma)
ax2 = fig.add_subplot(2, 2, 2)
# Extraer triangulo superior (sin diagonal) para evitar duplicados
upper_triangle = cosSimilarities_vgg16[np.triu_indices_from(cosSimilarities_vgg16, k=1)]
ax2.hist(upper_triangle, bins=80, color='#4CAF50', alpha=0.7, edgecolor='black', linewidth=0.5)
ax2.axvline(upper_triangle.mean(), color='red', linestyle='--', linewidth=2, label=f'Media: {upper_triangle.mean():.4f}')
ax2.axvline(np.median(upper_triangle), color='orange', linestyle='--', linewidth=2, label=f'Mediana: {np.median(upper_triangle):.4f}')
ax2.set_title('Distribucion de Similitudes entre Imagenes', fontsize=14, fontweight='bold')
ax2.set_xlabel('Score de Similitud Coseno', fontsize=11)
ax2.set_ylabel('Frecuencia', fontsize=11)
ax2.legend(fontsize=10)
ax2.grid(True, alpha=0.3)

# 3. Top 15 pares mas similares (excluyendo la misma imagen)
ax3 = fig.add_subplot(2, 2, 3)
# Encontrar los pares mas similares
n_images = len(files)
pairs = []
for i in range(n_images):
    for j in range(i+1, n_images):
        pairs.append((i, j, cosSimilarities_vgg16[i, j]))

# Ordenar por similitud descendente
pairs_sorted = sorted(pairs, key=lambda x: x[2], reverse=True)[:15]

# Graficar barras horizontales
pair_labels = [f'Img {p[0]} - Img {p[1]}' for p in pairs_sorted]
pair_scores = [p[2] for p in pairs_sorted]
colors_gradient = plt.cm.RdYlGn(np.linspace(0.9, 0.5, len(pair_scores)))

bars = ax3.barh(range(len(pair_labels)), pair_scores, color=colors_gradient, edgecolor='black', linewidth=0.5)
ax3.set_yticks(range(len(pair_labels)))
ax3.set_yticklabels(pair_labels, fontsize=9)
ax3.set_xlabel('Similitud Coseno', fontsize=11)
ax3.set_title('Top 15 Pares de Imagenes Mas Similares', fontsize=14, fontweight='bold')
ax3.invert_yaxis()
ax3.grid(True, axis='x', alpha=0.3)

# Añadir valores en las barras
for bar, score in zip(bars, pair_scores):
    ax3.text(bar.get_width() + 0.005, bar.get_y() + bar.get_height()/2,
             f'{score:.4f}', va='center', fontsize=8, fontweight='bold')

# 4. Estadisticas y boxplot
ax4 = fig.add_subplot(2, 2, 4)

# Boxplot con violin plot combinado
parts = ax4.violinplot([upper_triangle], positions=[1], showmeans=True, showmedians=True)
parts['bodies'][0].set_facecolor('#4CAF50')
parts['bodies'][0].set_alpha(0.7)

# Boxplot superpuesto
bp = ax4.boxplot([upper_triangle], positions=[1], widths=0.15, patch_artist=True)
bp['boxes'][0].set_facecolor('#2ecc71')
bp['boxes'][0].set_alpha(0.7)

ax4.set_ylabel('Similitud Coseno', fontsize=11)
ax4.set_title('Distribucion de Similitudes\n(Violin + Box Plot)', fontsize=14, fontweight='bold')
ax4.set_xticks([1])
ax4.set_xticklabels(['Todas las\ncomparaciones'], fontsize=11)
ax4.grid(True, axis='y', alpha=0.3)

# Añadir estadisticas como texto
stats_text = f"""
ESTADISTICAS GLOBALES (VGG16)
-----------------------------
Total imagenes: {n_images:,}
Total comparaciones: {len(upper_triangle):,}

Similitud minima: {upper_triangle.min():.4f}
Similitud maxima: {upper_triangle.max():.4f}
Media: {upper_triangle.mean():.4f}
Mediana: {np.median(upper_triangle):.4f}
Desv. Estandar: {upper_triangle.std():.4f}

Percentil 25: {np.percentile(upper_triangle, 25):.4f}
Percentil 75: {np.percentile(upper_triangle, 75):.4f}
Percentil 95: {np.percentile(upper_triangle, 95):.4f}
"""
ax4.text(1.6, 0.5, stats_text, transform=ax4.get_xaxis_transform(), fontsize=9,
         verticalalignment='center', fontfamily='monospace',
         bbox=dict(boxstyle='round', facecolor='#f8f9fa', alpha=0.9, edgecolor='gray'))

ax4.set_xlim(0.5, 2.8)

plt.suptitle('ANALISIS DE SIMILITUD COSENO ENTRE IMAGENES - VGG16',
             fontsize=18, fontweight='bold', y=1.02)
plt.tight_layout()
plt.show()

print(f"\nMatriz de similitud VGG16 calculada exitosamente!")
print(f"Dimensiones: {cosSimilarities_vgg16.shape[0]} x {cosSimilarities_vgg16.shape[1]} = {cosSimilarities_vgg16.shape[0]**2:,} comparaciones")

Modelo: VGG16
Imagen de referencia: idx = 1881
Archivo: images/3_2_045.png
============================================================

Matriz de similitud VGG16 calculada exitosamente!
Dimensiones: 2184 x 2184 = 4,769,856 comparaciones

# esta funcion recupera las imagenes mas similares dada una imagen entregada por el usuario (VGG16)
def retrieve_most_similar_products_vgg16(given_img):
    print("-----------------------------------------------------------------------")
    print("Modelo: VGG16")
    print("Producto escogido:")
    original = load_img(given_img, target_size=(imgs_model_width, imgs_model_height))
    plt.imshow(original)
    plt.show()
    print("-----------------------------------------------------------------------")
    print("Productos mas similares:")
    closest_imgs = cos_similarities_df_vgg16[given_img].sort_values(ascending=False)[1:nb_closest_images+1].index
    closest_imgs_scores = cos_similarities_df_vgg16[given_img].sort_values(ascending=False)[1:nb_closest_images+1]

    for i in range(0, len(closest_imgs)):
        original = load_img(closest_imgs[i], target_size=(imgs_model_width, imgs_model_height))
        plt.imshow(original)
        plt.show()
        print("Score de similaridad:", closest_imgs_scores[i])

# ============================================================================
# VISUALIZACION DE RECOMENDACIONES CON ANALISIS DE FEATURES - VGG16
# ============================================================================

query_image = files[idx]

print("Modelo: VGG16")
print(f"Indice seleccionado: {idx}")
print(f"Archivo: {query_image}")

# Obtener las imagenes mas similares y sus scores
closest_imgs = cos_similarities_df_vgg16[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
closest_imgs_paths = closest_imgs.index.tolist()
closest_imgs_scores = closest_imgs.values

# Obtener indices para extraer features
query_idx = files.index(query_image)
rec_indices = [files.index(p) for p in closest_imgs_paths]

# Colores para ranking
ranking_colors = ['#27ae60', '#2ecc71', '#f1c40f', '#e67e22', '#e74c3c']

# ============================================================================
# FIGURA 1: Imagen de Consulta
# ============================================================================
fig1, ax1 = plt.subplots(figsize=(6, 6), facecolor='#f8f9fa')
query_img = load_img(query_image, target_size=(imgs_model_width, imgs_model_height))
ax1.imshow(query_img)
ax1.set_title('IMAGEN DE CONSULTA - VGG16', fontsize=18, fontweight='bold', color='#2c3e50', pad=20)
ax1.axis('off')

# Borde elegante (verde para VGG16)
for spine in ax1.spines.values():
    spine.set_visible(True)
    spine.set_color('#4CAF50')
    spine.set_linewidth(5)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 2: Top 5 Recomendaciones
# ============================================================================
print("\n" + "="*70)
print("TOP 5 PRODUCTOS RECOMENDADOS - VGG16")
print("="*70 + "\n")

fig2, axes2 = plt.subplots(1, nb_closest_images, figsize=(18, 4), facecolor='#f8f9fa')

for i in range(nb_closest_images):
    ax = axes2[i]
    rec_img = load_img(closest_imgs_paths[i], target_size=(imgs_model_width, imgs_model_height))
    ax.imshow(rec_img)
    ax.axis('off')

    score = closest_imgs_scores[i]
    ax.set_title(f'#{i+1} | Similitud: {score:.4f}', fontsize=13, fontweight='bold',
                 color=ranking_colors[i], pad=15)

    # Borde con color de ranking
    for spine in ax.spines.values():
        spine.set_visible(True)
        spine.set_color(ranking_colors[i])
        spine.set_linewidth(4)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 3: Comparacion de Features (Barras)
# ============================================================================
print("\n" + "="*70)
print("ANALISIS COMPARATIVO DE FEATURES - VGG16")
print("="*70 + "\n")

fig3, ax3 = plt.subplots(figsize=(12, 5), facecolor='#f8f9fa')

bars = ax3.barh(range(nb_closest_images), closest_imgs_scores,
                color=ranking_colors, edgecolor='#2c3e50', height=0.6, linewidth=1.5)

ax3.set_yticks(range(nb_closest_images))
ax3.set_yticklabels([f'Recomendacion #{i+1}' for i in range(nb_closest_images)], fontsize=12, fontweight='bold')
ax3.set_xlabel('Score de Similitud Coseno', fontsize=13, fontweight='bold')
ax3.set_title('RANKING DE SIMILITUD - VGG16', fontsize=16, fontweight='bold', color='#2c3e50', pad=20)
ax3.set_xlim(0, 1)
ax3.invert_yaxis()
ax3.set_facecolor('#fafafa')
ax3.grid(True, axis='x', alpha=0.4, linestyle='--')
ax3.spines['top'].set_visible(False)
ax3.spines['right'].set_visible(False)

for bar, score in zip(bars, closest_imgs_scores):
    ax3.text(score + 0.02, bar.get_y() + bar.get_height()/2,
             f'{score:.4f}', va='center', fontsize=12, fontweight='bold')

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 4: Comparacion de Vectores de Features
# ============================================================================
fig4, axes4 = plt.subplots(2, 3, figsize=(16, 10), facecolor='#f8f9fa')

# Features de la imagen query
query_features = imgs_features_vgg16[query_idx]

# Plot 1: Distribucion de features de la imagen query
ax = axes4[0, 0]
ax.hist(query_features, bins=50, color='#4CAF50', alpha=0.7, edgecolor='black', linewidth=0.5)
ax.set_title('Distribucion de Features\n(Imagen Query - VGG16)', fontsize=12, fontweight='bold')
ax.set_xlabel('Valor del Feature')
ax.set_ylabel('Frecuencia')
ax.axvline(query_features.mean(), color='red', linestyle='--', linewidth=2, label=f'Media: {query_features.mean():.2f}')
ax.legend(fontsize=9)
ax.grid(True, alpha=0.3)

# Plot 2: Primeros 100 features comparados (Query vs Top 1)
ax = axes4[0, 1]
feature_range = range(100)
ax.plot(feature_range, query_features[:100], 'g-', alpha=0.7, linewidth=1.5, label='Query')
ax.plot(feature_range, imgs_features_vgg16[rec_indices[0]][:100], 'b--', alpha=0.7, linewidth=1.5, label='Top #1')
ax.set_title('Comparacion de Features (0-100)\nQuery vs Top #1', fontsize=12, fontweight='bold')
ax.set_xlabel('Indice del Feature')
ax.set_ylabel('Valor')
ax.legend(fontsize=10)
ax.grid(True, alpha=0.3)

# Plot 3: Heatmap de correlacion entre query y recomendaciones
ax = axes4[0, 2]
correlation_matrix = np.zeros((6, 6))
all_features = [query_features] + [imgs_features_vgg16[rec_idx] for rec_idx in rec_indices]
labels = ['Query'] + [f'#{i+1}' for i in range(nb_closest_images)]

for i in range(6):
    for j in range(6):
        correlation_matrix[i, j] = cosine_similarity([all_features[i]], [all_features[j]])[0][0]

im = ax.imshow(correlation_matrix, cmap='RdYlGn', vmin=0.5, vmax=1)
ax.set_xticks(range(6))
ax.set_yticks(range(6))
ax.set_xticklabels(labels, fontsize=10, fontweight='bold')
ax.set_yticklabels(labels, fontsize=10, fontweight='bold')
ax.set_title('Matriz de Similitud\nQuery vs Recomendaciones', fontsize=12, fontweight='bold')
plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)

# Añadir valores en el heatmap
for i in range(6):
    for j in range(6):
        ax.text(j, i, f'{correlation_matrix[i,j]:.2f}', ha='center', va='center',
                fontsize=9, fontweight='bold', color='white' if correlation_matrix[i,j] > 0.75 else 'black')

# Plot 4: Activaciones mas altas (Top features)
ax = axes4[1, 0]
top_k = 20
top_indices = np.argsort(query_features)[-top_k:][::-1]
top_values = query_features[top_indices]

colors_bar = plt.cm.Greens(np.linspace(0.4, 0.9, top_k))
ax.barh(range(top_k), top_values, color=colors_bar, edgecolor='black', linewidth=0.5)
ax.set_yticks(range(top_k))
ax.set_yticklabels([f'F-{idx}' for idx in top_indices], fontsize=8)
ax.set_xlabel('Valor de Activacion')
ax.set_title(f'Top {top_k} Features con Mayor Activacion\n(Imagen Query - VGG16)', fontsize=12, fontweight='bold')
ax.invert_yaxis()
ax.grid(True, axis='x', alpha=0.3)

# Plot 5: Diferencia de features entre Query y cada recomendacion
ax = axes4[1, 1]
differences = []
for i, rec_idx in enumerate(rec_indices):
    diff = np.mean(np.abs(query_features - imgs_features_vgg16[rec_idx]))
    differences.append(diff)

bars = ax.bar(range(nb_closest_images), differences, color=ranking_colors, edgecolor='black', linewidth=1.5)
ax.set_xticks(range(nb_closest_images))
ax.set_xticklabels([f'#{i+1}' for i in range(nb_closest_images)], fontsize=11, fontweight='bold')
ax.set_ylabel('Diferencia Promedio Absoluta')
ax.set_title('Diferencia de Features\nQuery vs Recomendaciones', fontsize=12, fontweight='bold')
ax.grid(True, axis='y', alpha=0.3)

# Posicionar texto justo encima de cada barra
for i, (bar, diff) in enumerate(zip(bars, differences)):
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{diff:.2f}', ha='center', va='bottom', fontsize=10, fontweight='bold')

# Plot 6: Estadisticas resumen
ax = axes4[1, 2]
ax.axis('off')

stats_text = f"""
ESTADISTICAS DE FEATURES (VGG16)
{'='*40}

IMAGEN QUERY:
  - Dimension del vector: {len(query_features):,}
  - Media: {query_features.mean():.4f}
  - Std: {query_features.std():.4f}
  - Min: {query_features.min():.4f}
  - Max: {query_features.max():.4f}
  - Features activos (>0): {np.sum(query_features > 0):,}

SIMILITUDES:
  - Mejor match: #{1} ({closest_imgs_scores[0]:.4f})
  - Peor match: #{nb_closest_images} ({closest_imgs_scores[-1]:.4f})
  - Rango: {closest_imgs_scores[0] - closest_imgs_scores[-1]:.4f}

DIFERENCIAS PROMEDIO:
  - Menor diferencia: #{np.argmin(differences)+1} ({min(differences):.2f})
  - Mayor diferencia: #{np.argmax(differences)+1} ({max(differences):.2f})
"""

ax.text(0.1, 0.95, stats_text, transform=ax.transAxes, fontsize=11,
        verticalalignment='top', fontfamily='monospace',
        bbox=dict(boxstyle='round', facecolor='#e8f8e8', alpha=0.9, edgecolor='#4CAF50', linewidth=2))

plt.suptitle('ANALISIS DETALLADO DE FEATURES CNN - VGG16', fontsize=18, fontweight='bold', color='#2c3e50', y=1.02)
plt.tight_layout()
plt.show()

# ============================================================================
# RESUMEN TABULAR
# ============================================================================
print("\n")
print("+" + "="*62 + "+")
print("|{:^62}|".format("RESUMEN DE RECOMENDACIONES - VGG16"))
print("+" + "="*62 + "+")
print("| {:<10} | {:<15} | {:<30} |".format("Ranking", "Similitud", "Archivo"))
print("+" + "-"*62 + "+")
for i, (path, score) in enumerate(zip(closest_imgs_paths, closest_imgs_scores)):
    filename = path.split('/')[-1]
    print("| {:<10} | {:<15.4f} | {:<30} |".format(f"#{i+1}", score, filename))
print("+" + "="*62 + "+")

Modelo: VGG16
Indice seleccionado: 1881
Archivo: images/3_2_045.png

======================================================================
TOP 5 PRODUCTOS RECOMENDADOS - VGG16
======================================================================

======================================================================
ANALISIS COMPARATIVO DE FEATURES - VGG16
======================================================================


+==============================================================+
|              RESUMEN DE RECOMENDACIONES - VGG16              |
+==============================================================+
| Ranking    | Similitud       | Archivo                        |
+--------------------------------------------------------------+
| #1         | 0.6578          | 5_2_059.png                    |
| #2         | 0.6397          | 3_2_066.png                    |
| #3         | 0.6396          | 6_2_033.png                    |
| #4         | 0.6349          | 3_2_043.png                    |
| #5         | 0.6058          | 5_2_044.png                    |
+==============================================================+

# ============================================================================
# COMPARACION FINAL: VGG16 vs VGG19
# ============================================================================

print("="*70)
print("COMPARACION DE MODELOS: VGG16 vs VGG19")
print("="*70)
print(f"\nImagen de consulta: idx = {idx}")
print(f"Archivo: {files[idx]}\n")

# Obtener recomendaciones de ambos modelos
query_image = files[idx]

# VGG16
closest_vgg16 = cos_similarities_df_vgg16[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
paths_vgg16 = closest_vgg16.index.tolist()
scores_vgg16 = closest_vgg16.values

# VGG19
closest_vgg19 = cos_similarities_df[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
paths_vgg19 = closest_vgg19.index.tolist()
scores_vgg19 = closest_vgg19.values

# ============================================================================
# FIGURA 1: Imagen Query + Recomendaciones lado a lado
# ============================================================================
fig1 = plt.figure(figsize=(20, 14), facecolor='white')

# Titulo principal
fig1.suptitle('COMPARACION DE RECOMENDACIONES',
              fontsize=24, fontweight='bold', color='#2c3e50', y=0.97)
fig1.text(0.5, 0.93, 'VGG16 vs VGG19', fontsize=18, color='#7f8c8d', ha='center', style='italic')

# --- IMAGEN DE CONSULTA ---
ax_query = fig1.add_axes([0.35, 0.72, 0.30, 0.18])
query_img = load_img(query_image, target_size=(imgs_model_width, imgs_model_height))
ax_query.imshow(query_img)
ax_query.set_title('IMAGEN DE CONSULTA', fontsize=16, fontweight='bold', color='#2c3e50', pad=15)
ax_query.axis('off')
for spine in ax_query.spines.values():
    spine.set_visible(True)
    spine.set_color('#9b59b6')
    spine.set_linewidth(5)

# --- SECCION VGG16 ---
# Fondo verde claro para VGG16
ax_bg_vgg16 = fig1.add_axes([0.01, 0.36, 0.98, 0.32])
ax_bg_vgg16.set_facecolor('#e8f5e9')
ax_bg_vgg16.axis('off')

# Etiqueta VGG16
fig1.text(0.02, 0.65, 'VGG16', fontsize=22, fontweight='bold', color='white', ha='left',
          bbox=dict(boxstyle='round,pad=0.4', facecolor='#4CAF50', edgecolor='#2e7d32', linewidth=3))

# Recomendaciones VGG16
for i in range(nb_closest_images):
    ax = fig1.add_axes([0.02 + i*0.19, 0.38, 0.17, 0.26])
    rec_img = load_img(paths_vgg16[i], target_size=(imgs_model_width, imgs_model_height))
    ax.imshow(rec_img)
    ax.set_title(f'#{i+1} | Score: {scores_vgg16[i]:.4f}', fontsize=12, fontweight='bold', color='#2e7d32', pad=8)
    ax.axis('off')
    for spine in ax.spines.values():
        spine.set_visible(True)
        spine.set_color('#4CAF50')
        spine.set_linewidth(3)

# --- LINEA DIVISORIA ---
fig1.add_axes([0.05, 0.34, 0.90, 0.005]).set_facecolor('#bdc3c7')
fig1.axes[-1].axis('off')

# --- SECCION VGG19 ---
# Fondo azul claro para VGG19
ax_bg_vgg19 = fig1.add_axes([0.01, 0.02, 0.98, 0.32])
ax_bg_vgg19.set_facecolor('#e3f2fd')
ax_bg_vgg19.axis('off')

# Etiqueta VGG19
fig1.text(0.02, 0.31, 'VGG19', fontsize=22, fontweight='bold', color='white', ha='left',
          bbox=dict(boxstyle='round,pad=0.4', facecolor='#2196F3', edgecolor='#1565c0', linewidth=3))

# Recomendaciones VGG19
for i in range(nb_closest_images):
    ax = fig1.add_axes([0.02 + i*0.19, 0.04, 0.17, 0.26])
    rec_img = load_img(paths_vgg19[i], target_size=(imgs_model_width, imgs_model_height))
    ax.imshow(rec_img)
    ax.set_title(f'#{i+1} | Score: {scores_vgg19[i]:.4f}', fontsize=12, fontweight='bold', color='#1565c0', pad=8)
    ax.axis('off')
    for spine in ax.spines.values():
        spine.set_visible(True)
        spine.set_color('#2196F3')
        spine.set_linewidth(3)

plt.show()

# ============================================================================
# FIGURA 2: Comparacion de Scores
# ============================================================================
fig2, axes2 = plt.subplots(1, 3, figsize=(18, 5), facecolor='#f8f9fa')

# Plot 1: Barras comparativas de scores
ax = axes2[0]
x = np.arange(nb_closest_images)
width = 0.35

bars1 = ax.bar(x - width/2, scores_vgg16, width, label='VGG16', color='#4CAF50', edgecolor='black', linewidth=1)
bars2 = ax.bar(x + width/2, scores_vgg19, width, label='VGG19', color='#2196F3', edgecolor='black', linewidth=1)

ax.set_xlabel('Ranking de Recomendacion', fontsize=12, fontweight='bold')
ax.set_ylabel('Score de Similitud', fontsize=12, fontweight='bold')
ax.set_title('Comparacion de Scores por Ranking', fontsize=14, fontweight='bold')
ax.set_xticks(x)
ax.set_xticklabels([f'#{i+1}' for i in range(nb_closest_images)], fontsize=11, fontweight='bold')
ax.legend(fontsize=11)
ax.grid(True, axis='y', alpha=0.3)
ax.set_ylim(0, 1)

# Añadir valores
for bar in bars1:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.3f}', ha='center', va='bottom', fontsize=8, fontweight='bold')
for bar in bars2:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.3f}', ha='center', va='bottom', fontsize=8, fontweight='bold')

# Plot 2: Diferencia de scores
ax = axes2[1]
diff_scores = scores_vgg16 - scores_vgg19
colors_diff = ['#4CAF50' if d > 0 else '#2196F3' for d in diff_scores]

bars = ax.bar(range(nb_closest_images), diff_scores, color=colors_diff, edgecolor='black', linewidth=1.5)
ax.axhline(y=0, color='black', linestyle='-', linewidth=1)
ax.set_xlabel('Ranking de Recomendacion', fontsize=12, fontweight='bold')
ax.set_ylabel('Diferencia (VGG16 - VGG19)', fontsize=12, fontweight='bold')
ax.set_title('Diferencia de Scores entre Modelos', fontsize=14, fontweight='bold')
ax.set_xticks(range(nb_closest_images))
ax.set_xticklabels([f'#{i+1}' for i in range(nb_closest_images)], fontsize=11, fontweight='bold')
ax.grid(True, axis='y', alpha=0.3)

for i, (bar, diff) in enumerate(zip(bars, diff_scores)):
    y_pos = bar.get_height() + 0.005 if diff >= 0 else bar.get_height() - 0.015
    ax.text(bar.get_x() + bar.get_width()/2, y_pos,
            f'{diff:.4f}', ha='center', va='bottom' if diff >= 0 else 'top',
            fontsize=9, fontweight='bold')

# Plot 3: Distribucion global de similitudes
ax = axes2[2]
upper_vgg16 = cosSimilarities_vgg16[np.triu_indices_from(cosSimilarities_vgg16, k=1)]
upper_vgg19 = cosSimilarities[np.triu_indices_from(cosSimilarities, k=1)]

ax.hist(upper_vgg16, bins=50, alpha=0.6, color='#4CAF50', label=f'VGG16 (mean={upper_vgg16.mean():.4f})', edgecolor='black', linewidth=0.5)
ax.hist(upper_vgg19, bins=50, alpha=0.6, color='#2196F3', label=f'VGG19 (mean={upper_vgg19.mean():.4f})', edgecolor='black', linewidth=0.5)
ax.axvline(upper_vgg16.mean(), color='#2e7d32', linestyle='--', linewidth=2)
ax.axvline(upper_vgg19.mean(), color='#1565c0', linestyle='--', linewidth=2)
ax.set_xlabel('Score de Similitud Coseno', fontsize=12, fontweight='bold')
ax.set_ylabel('Frecuencia', fontsize=12, fontweight='bold')
ax.set_title('Distribucion Global de Similitudes', fontsize=14, fontweight='bold')
ax.legend(fontsize=10)
ax.grid(True, alpha=0.3)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 3: Comparacion de Parametros y Arquitectura
# ============================================================================
fig3, axes3 = plt.subplots(1, 2, figsize=(14, 5), facecolor='#f8f9fa')

# Plot 1: Parametros
ax = axes3[0]
params_vgg16 = feat_extractor_vgg16.count_params()
params_vgg19 = feat_extractor.count_params()

bars = ax.bar(['VGG16', 'VGG19'], [params_vgg16, params_vgg19],
              color=['#4CAF50', '#2196F3'], edgecolor='black', linewidth=2)
ax.set_ylabel('Numero de Parametros', fontsize=12, fontweight='bold')
ax.set_title('Comparacion de Parametros Entrenables', fontsize=14, fontweight='bold')
ax.yaxis.set_major_formatter(plt.FuncFormatter(lambda x, p: f'{int(x/1e6)}M'))
ax.grid(True, axis='y', alpha=0.3)

for bar in bars:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 1e6,
            f'{bar.get_height():,.0f}', ha='center', va='bottom', fontsize=10, fontweight='bold')

# Plot 2: Metricas comparativas
ax = axes3[1]
metrics = ['Media Similitud', 'Std Similitud', 'Score Top-1', 'Score Top-5']
values_vgg16 = [upper_vgg16.mean(), upper_vgg16.std(), scores_vgg16[0], scores_vgg16[-1]]
values_vgg19 = [upper_vgg19.mean(), upper_vgg19.std(), scores_vgg19[0], scores_vgg19[-1]]

x = np.arange(len(metrics))
width = 0.35

bars1 = ax.bar(x - width/2, values_vgg16, width, label='VGG16', color='#4CAF50', edgecolor='black')
bars2 = ax.bar(x + width/2, values_vgg19, width, label='VGG19', color='#2196F3', edgecolor='black')

ax.set_ylabel('Valor', fontsize=12, fontweight='bold')
ax.set_title('Metricas Comparativas', fontsize=14, fontweight='bold')
ax.set_xticks(x)
ax.set_xticklabels(metrics, fontsize=10, fontweight='bold')
ax.legend(fontsize=11)
ax.grid(True, axis='y', alpha=0.3)

for bar in bars1:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.4f}', ha='center', va='bottom', fontsize=8, fontweight='bold', rotation=45)
for bar in bars2:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.4f}', ha='center', va='bottom', fontsize=8, fontweight='bold', rotation=45)

plt.tight_layout()
plt.show()

# ============================================================================
# FIGURA 4: Coincidencias en recomendaciones
# ============================================================================
# Verificar cuantas recomendaciones coinciden
coincidencias = set(paths_vgg16) & set(paths_vgg19)
solo_vgg16 = set(paths_vgg16) - set(paths_vgg19)
solo_vgg19 = set(paths_vgg19) - set(paths_vgg16)

print("\n" + "="*70)
print("ANALISIS DE COINCIDENCIAS EN RECOMENDACIONES")
print("="*70)
print(f"\nImagenes recomendadas por AMBOS modelos: {len(coincidencias)}")
print(f"Solo recomendadas por VGG16: {len(solo_vgg16)}")
print(f"Solo recomendadas por VGG19: {len(solo_vgg19)}")
print(f"Porcentaje de coincidencia: {len(coincidencias)/nb_closest_images*100:.1f}%")

if len(coincidencias) > 0:
    fig4, axes4 = plt.subplots(1, len(coincidencias), figsize=(4*len(coincidencias), 4), facecolor='#f8f9fa')
    if len(coincidencias) == 1:
        axes4 = [axes4]

    for i, path in enumerate(coincidencias):
        ax = axes4[i]
        img = load_img(path, target_size=(imgs_model_width, imgs_model_height))
        ax.imshow(img)

        # Obtener scores de ambos modelos
        score_16 = cos_similarities_df_vgg16[query_image][path]
        score_19 = cos_similarities_df[query_image][path]

        ax.set_title(f'VGG16: {score_16:.4f}\nVGG19: {score_19:.4f}', fontsize=11, fontweight='bold')
        ax.axis('off')
        for spine in ax.spines.values():
            spine.set_visible(True)
            spine.set_color('#9b59b6')
            spine.set_linewidth(3)

    plt.suptitle('IMAGENES RECOMENDADAS POR AMBOS MODELOS', fontsize=16, fontweight='bold', y=1.05)
    plt.tight_layout()
    plt.show()

# ============================================================================
# TABLA COMPARATIVA FINAL
# ============================================================================
print("\n" + "+"*80)
print("|{:^78}|".format("TABLA COMPARATIVA DE RECOMENDACIONES"))
print("+"*80)
print("|{:^10}|{:^32}|{:^32}|".format("Ranking", "VGG16", "VGG19"))
print("|{:^10}|{:^15}|{:^15}|{:^15}|{:^15}|".format("", "Score", "Archivo", "Score", "Archivo"))
print("+"*80)

for i in range(nb_closest_images):
    file16 = paths_vgg16[i].split('/')[-1][:12]
    file19 = paths_vgg19[i].split('/')[-1][:12]
    match = " *" if paths_vgg16[i] == paths_vgg19[i] else ""
    print("|{:^10}|{:^15.4f}|{:^15}|{:^15.4f}|{:^15}|{}".format(
        f"#{i+1}", scores_vgg16[i], file16, scores_vgg19[i], file19, match))

print("+"*80)
print("| * = Misma imagen recomendada por ambos modelos" + " "*30 + "|")
print("+"*80)

# ============================================================================
# RESUMEN ESTADISTICO FINAL
# ============================================================================
print("\n" + "="*70)
print("RESUMEN ESTADISTICO FINAL")
print("="*70)

print(f"""
{'Metrica':<35} {'VGG16':>15} {'VGG19':>15}
{'-'*65}
{'Parametros totales':<35} {params_vgg16:>15,} {params_vgg19:>15,}
{'Similitud media global':<35} {upper_vgg16.mean():>15.4f} {upper_vgg19.mean():>15.4f}
{'Similitud std global':<35} {upper_vgg16.std():>15.4f} {upper_vgg19.std():>15.4f}
{'Score Top-1':<35} {scores_vgg16[0]:>15.4f} {scores_vgg19[0]:>15.4f}
{'Score Top-5':<35} {scores_vgg16[-1]:>15.4f} {scores_vgg19[-1]:>15.4f}
{'Score promedio Top-5':<35} {scores_vgg16.mean():>15.4f} {scores_vgg19.mean():>15.4f}
{'Coincidencias en Top-5':<35} {len(coincidencias):>15} {len(coincidencias):>15}
{'-'*65}
""")

# Conclusion
mejor_top1 = "VGG16" if scores_vgg16[0] > scores_vgg19[0] else "VGG19"
mejor_promedio = "VGG16" if scores_vgg16.mean() > scores_vgg19.mean() else "VGG19"

print("CONCLUSION:")
print(f"  - Mejor score Top-1: {mejor_top1}")
print(f"  - Mejor promedio Top-5: {mejor_promedio}")
print(f"  - VGG19 tiene {params_vgg19 - params_vgg16:,} parametros mas ({(params_vgg19-params_vgg16)/params_vgg16*100:.1f}%)")
print("="*70)

======================================================================
COMPARACION DE MODELOS: VGG16 vs VGG19
======================================================================

Imagen de consulta: idx = 1881
Archivo: images/3_2_045.png

======================================================================
ANALISIS DE COINCIDENCIAS EN RECOMENDACIONES
======================================================================

Imagenes recomendadas por AMBOS modelos: 2
Solo recomendadas por VGG16: 3
Solo recomendadas por VGG19: 3
Porcentaje de coincidencia: 40.0%

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
|                     TABLA COMPARATIVA DE RECOMENDACIONES                     |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
| Ranking  |             VGG16              |             VGG19              |
|          |     Score     |    Archivo    |     Score     |    Archivo    |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
|    #1    |    0.6578     |  5_2_059.png  |    0.6318     |  3_2_066.png  |
|    #2    |    0.6397     |  3_2_066.png  |    0.6237     |  4_2_030.png  |
|    #3    |    0.6396     |  6_2_033.png  |    0.5985     |  4_2_029.png  |
|    #4    |    0.6349     |  3_2_043.png  |    0.5978     |  0_2_006.png  |
|    #5    |    0.6058     |  5_2_044.png  |    0.5943     |  5_2_059.png  |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
| * = Misma imagen recomendada por ambos modelos                              |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

======================================================================
RESUMEN ESTADISTICO FINAL
======================================================================

Metrica                                       VGG16           VGG19
-----------------------------------------------------------------
Parametros totales                      134,260,544     139,570,240
Similitud media global                       0.3109          0.3019
Similitud std global                         0.1116          0.1112
Score Top-1                                  0.6578          0.6318
Score Top-5                                  0.6058          0.5943
Score promedio Top-5                         0.6356          0.6092
Coincidencias en Top-5                            2               2
-----------------------------------------------------------------

CONCLUSION:
  - Mejor score Top-1: VGG16
  - Mejor promedio Top-5: VGG16
  - VGG19 tiene 5,309,696 parametros mas (4.0%)
======================================================================

# ============================================================================
# FUNCION AUXILIAR PARA COMPARAR RECOMENDACIONES VGG16 vs VGG19
# ============================================================================

def comparar_recomendaciones(idx_imagen, titulo_ejemplo):
    """
    Genera una visualizacion comparativa de recomendaciones entre VGG16 y VGG19
    para una imagen de consulta especifica.
    """
    query_image = files[idx_imagen]

    # Obtener recomendaciones VGG16
    closest_vgg16 = cos_similarities_df_vgg16[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
    paths_vgg16 = closest_vgg16.index.tolist()
    scores_vgg16 = closest_vgg16.values

    # Obtener recomendaciones VGG19
    closest_vgg19 = cos_similarities_df[query_image].sort_values(ascending=False)[1:nb_closest_images+1]
    paths_vgg19 = closest_vgg19.index.tolist()
    scores_vgg19 = closest_vgg19.values

    # Calcular coincidencias
    coincidencias = set(paths_vgg16) & set(paths_vgg19)

    # --- FIGURA COMPARATIVA ---
    fig = plt.figure(figsize=(20, 14), facecolor='white')

    # Titulo principal
    fig.suptitle(f'{titulo_ejemplo}', fontsize=22, fontweight='bold', color='#2c3e50', y=0.98)
    fig.text(0.5, 0.93, f'Imagen de consulta: idx = {idx_imagen} | Coincidencias: {len(coincidencias)}/5',
             fontsize=14, color='#7f8c8d', ha='center', style='italic')

    # --- IMAGEN DE CONSULTA ---
    ax_query = fig.add_axes([0.35, 0.72, 0.30, 0.18])
    query_img = load_img(query_image, target_size=(imgs_model_width, imgs_model_height))
    ax_query.imshow(query_img)
    ax_query.set_title('IMAGEN DE CONSULTA', fontsize=16, fontweight='bold', color='#2c3e50', pad=15)
    ax_query.axis('off')
    for spine in ax_query.spines.values():
        spine.set_visible(True)
        spine.set_color('#9b59b6')
        spine.set_linewidth(5)

    # --- SECCION VGG16 ---
    ax_bg_vgg16 = fig.add_axes([0.01, 0.36, 0.98, 0.32])
    ax_bg_vgg16.set_facecolor('#e8f5e9')
    ax_bg_vgg16.axis('off')

    fig.text(0.02, 0.65, 'VGG16', fontsize=20, fontweight='bold', color='white', ha='left',
             bbox=dict(boxstyle='round,pad=0.4', facecolor='#4CAF50', edgecolor='#2e7d32', linewidth=3))

    for i in range(nb_closest_images):
        ax = fig.add_axes([0.02 + i*0.19, 0.38, 0.17, 0.26])
        rec_img = load_img(paths_vgg16[i], target_size=(imgs_model_width, imgs_model_height))
        ax.imshow(rec_img)

        # Marcar si coincide con VGG19
        marca = " *" if paths_vgg16[i] in coincidencias else ""
        ax.set_title(f'#{i+1} | {scores_vgg16[i]:.4f}{marca}', fontsize=11, fontweight='bold', color='#2e7d32', pad=8)
        ax.axis('off')

        # Borde dorado si coincide
        color_borde = '#FFD700' if paths_vgg16[i] in coincidencias else '#4CAF50'
        for spine in ax.spines.values():
            spine.set_visible(True)
            spine.set_color(color_borde)
            spine.set_linewidth(4 if paths_vgg16[i] in coincidencias else 3)

    # --- LINEA DIVISORIA ---
    ax_linea = fig.add_axes([0.05, 0.34, 0.90, 0.005])
    ax_linea.set_facecolor('#bdc3c7')
    ax_linea.axis('off')

    # --- SECCION VGG19 ---
    ax_bg_vgg19 = fig.add_axes([0.01, 0.02, 0.98, 0.32])
    ax_bg_vgg19.set_facecolor('#e3f2fd')
    ax_bg_vgg19.axis('off')

    fig.text(0.02, 0.31, 'VGG19', fontsize=20, fontweight='bold', color='white', ha='left',
             bbox=dict(boxstyle='round,pad=0.4', facecolor='#2196F3', edgecolor='#1565c0', linewidth=3))

    for i in range(nb_closest_images):
        ax = fig.add_axes([0.02 + i*0.19, 0.04, 0.17, 0.26])
        rec_img = load_img(paths_vgg19[i], target_size=(imgs_model_width, imgs_model_height))
        ax.imshow(rec_img)

        # Marcar si coincide con VGG16
        marca = " *" if paths_vgg19[i] in coincidencias else ""
        ax.set_title(f'#{i+1} | {scores_vgg19[i]:.4f}{marca}', fontsize=11, fontweight='bold', color='#1565c0', pad=8)
        ax.axis('off')

        # Borde dorado si coincide
        color_borde = '#FFD700' if paths_vgg19[i] in coincidencias else '#2196F3'
        for spine in ax.spines.values():
            spine.set_visible(True)
            spine.set_color(color_borde)
            spine.set_linewidth(4 if paths_vgg19[i] in coincidencias else 3)

    # Leyenda
    fig.text(0.98, 0.35, '* = Recomendada por ambos modelos (borde dorado)',
             fontsize=10, color='#7f8c8d', ha='right', style='italic')

    plt.show()

    # --- TABLA RESUMEN ---
    print("\n" + "-"*70)
    print(f"RESUMEN {titulo_ejemplo}")
    print("-"*70)
    print(f"{'Ranking':<10}{'VGG16 Score':<15}{'VGG19 Score':<15}{'Diferencia':<15}{'Coincide'}")
    print("-"*70)

    for i in range(nb_closest_images):
        diff = scores_vgg16[i] - scores_vgg19[i]
        coincide = "SI" if paths_vgg16[i] == paths_vgg19[i] else "NO"
        print(f"#{i+1:<9}{scores_vgg16[i]:<15.4f}{scores_vgg19[i]:<15.4f}{diff:<+15.4f}{coincide}")

    print("-"*70)
    print(f"Promedio: {scores_vgg16.mean():<15.4f}{scores_vgg19.mean():<15.4f}{scores_vgg16.mean()-scores_vgg19.mean():<+15.4f}")
    print(f"Coincidencias totales: {len(coincidencias)}/5 ({len(coincidencias)/5*100:.0f}%)")
    print("-"*70 + "\n")

    return scores_vgg16, scores_vgg19, len(coincidencias)

# ============================================================================
# EJEMPLO 1: idx = 952
# ============================================================================

scores_ej1_vgg16, scores_ej1_vgg19, coinc_ej1 = comparar_recomendaciones(952, "EJEMPLO 1: Busqueda de Imagenes Similares")

----------------------------------------------------------------------
RESUMEN EJEMPLO 1: Busqueda de Imagenes Similares
----------------------------------------------------------------------
Ranking   VGG16 Score    VGG19 Score    Diferencia     Coincide
----------------------------------------------------------------------
#1        0.6601         0.7389         -0.0788        NO
#2        0.6575         0.7060         -0.0486        NO
#3        0.6560         0.7031         -0.0471        NO
#4        0.6543         0.6852         -0.0309        NO
#5        0.6435         0.6849         -0.0414        NO
----------------------------------------------------------------------
Promedio: 0.6543         0.7036         -0.0494        
Coincidencias totales: 1/5 (20%)
----------------------------------------------------------------------

# ============================================================================
# EJEMPLO 2: idx = 396
# ============================================================================

scores_ej2_vgg16, scores_ej2_vgg19, coinc_ej2 = comparar_recomendaciones(396, "EJEMPLO 2: Busqueda de Imagenes Similares")

----------------------------------------------------------------------
RESUMEN EJEMPLO 2: Busqueda de Imagenes Similares
----------------------------------------------------------------------
Ranking   VGG16 Score    VGG19 Score    Diferencia     Coincide
----------------------------------------------------------------------
#1        0.8655         0.8434         +0.0221        SI
#2        0.7473         0.6815         +0.0658        SI
#3        0.7278         0.6808         +0.0470        NO
#4        0.7131         0.6780         +0.0352        NO
#5        0.7069         0.6764         +0.0305        NO
----------------------------------------------------------------------
Promedio: 0.7521         0.7120         +0.0401        
Coincidencias totales: 2/5 (40%)
----------------------------------------------------------------------

# ============================================================================
# EJEMPLO 3: idx = 26
# ============================================================================

scores_ej3_vgg16, scores_ej3_vgg19, coinc_ej3 = comparar_recomendaciones(26, "EJEMPLO 3: Busqueda de Imagenes Similares")

----------------------------------------------------------------------
RESUMEN EJEMPLO 3: Busqueda de Imagenes Similares
----------------------------------------------------------------------
Ranking   VGG16 Score    VGG19 Score    Diferencia     Coincide
----------------------------------------------------------------------
#1        0.8228         0.8092         +0.0137        SI
#2        0.7980         0.7984         -0.0004        NO
#3        0.7870         0.7945         -0.0075        NO
#4        0.7717         0.7705         +0.0012        NO
#5        0.7663         0.7440         +0.0222        NO
----------------------------------------------------------------------
Promedio: 0.7891         0.7833         +0.0058        
Coincidencias totales: 5/5 (100%)
----------------------------------------------------------------------

# ============================================================================
# RESUMEN FINAL: COMPARACION DE LOS 3 EJEMPLOS
# ============================================================================

print("="*70)
print("RESUMEN FINAL - PREGUNTA 1: BUSQUEDA DE IMAGENES SIMILARES")
print("="*70)

# Datos de los 3 ejemplos
ejemplos = ['Ejemplo 1 (idx=952)', 'Ejemplo 2 (idx=396)', 'Ejemplo 3 (idx=26)']
promedios_vgg16 = [scores_ej1_vgg16.mean(), scores_ej2_vgg16.mean(), scores_ej3_vgg16.mean()]
promedios_vgg19 = [scores_ej1_vgg19.mean(), scores_ej2_vgg19.mean(), scores_ej3_vgg19.mean()]
coincidencias_totales = [coinc_ej1, coinc_ej2, coinc_ej3]

# --- FIGURA RESUMEN ---
fig, axes = plt.subplots(1, 3, figsize=(18, 5), facecolor='#f8f9fa')

# Plot 1: Scores promedio por ejemplo
ax = axes[0]
x = np.arange(len(ejemplos))
width = 0.35

bars1 = ax.bar(x - width/2, promedios_vgg16, width, label='VGG16', color='#4CAF50', edgecolor='black')
bars2 = ax.bar(x + width/2, promedios_vgg19, width, label='VGG19', color='#2196F3', edgecolor='black')

ax.set_ylabel('Score Promedio Top-5', fontsize=12, fontweight='bold')
ax.set_title('Scores Promedio por Ejemplo', fontsize=14, fontweight='bold')
ax.set_xticks(x)
ax.set_xticklabels(['Ej. 1', 'Ej. 2', 'Ej. 3'], fontsize=11, fontweight='bold')
ax.legend(fontsize=10)
ax.grid(True, axis='y', alpha=0.3)
ax.set_ylim(0, 1)

for bar in bars1:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.3f}', ha='center', va='bottom', fontsize=9, fontweight='bold')
for bar in bars2:
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.01,
            f'{bar.get_height():.3f}', ha='center', va='bottom', fontsize=9, fontweight='bold')

# Plot 2: Coincidencias por ejemplo
ax = axes[1]
colors_coinc = ['#FFD700' if c >= 3 else '#e74c3c' if c <= 1 else '#f39c12' for c in coincidencias_totales]
bars = ax.bar(['Ej. 1', 'Ej. 2', 'Ej. 3'], coincidencias_totales, color=colors_coinc, edgecolor='black', linewidth=2)

ax.set_ylabel('Numero de Coincidencias', fontsize=12, fontweight='bold')
ax.set_title('Coincidencias entre Modelos', fontsize=14, fontweight='bold')
ax.set_ylim(0, 5)
ax.axhline(y=2.5, color='gray', linestyle='--', linewidth=1, alpha=0.5)
ax.grid(True, axis='y', alpha=0.3)

for bar, coinc in zip(bars, coincidencias_totales):
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 0.1,
            f'{coinc}/5', ha='center', va='bottom', fontsize=12, fontweight='bold')

# Plot 3: Diferencia VGG16 - VGG19
ax = axes[2]
diferencias = [p16 - p19 for p16, p19 in zip(promedios_vgg16, promedios_vgg19)]
colors_diff = ['#4CAF50' if d > 0 else '#2196F3' for d in diferencias]

bars = ax.bar(['Ej. 1', 'Ej. 2', 'Ej. 3'], diferencias, color=colors_diff, edgecolor='black', linewidth=2)
ax.axhline(y=0, color='black', linestyle='-', linewidth=1)
ax.set_ylabel('Diferencia (VGG16 - VGG19)', fontsize=12, fontweight='bold')
ax.set_title('Ventaja por Modelo', fontsize=14, fontweight='bold')
ax.grid(True, axis='y', alpha=0.3)

for bar, diff in zip(bars, diferencias):
    y_pos = bar.get_height() + 0.002 if diff >= 0 else bar.get_height() - 0.008
    modelo = 'VGG16' if diff > 0 else 'VGG19'
    ax.text(bar.get_x() + bar.get_width()/2, y_pos,
            f'{modelo}\n{abs(diff):.4f}', ha='center', va='bottom' if diff >= 0 else 'top',
            fontsize=9, fontweight='bold')

plt.tight_layout()
plt.show()

# --- TABLA FINAL ---
print("\n" + "="*70)
print(f"{'Ejemplo':<25}{'VGG16 Prom.':<15}{'VGG19 Prom.':<15}{'Mejor':<12}{'Coincid.'}")
print("="*70)

for i, ej in enumerate(ejemplos):
    mejor = "VGG16" if promedios_vgg16[i] > promedios_vgg19[i] else "VGG19"
    print(f"{ej:<25}{promedios_vgg16[i]:<15.4f}{promedios_vgg19[i]:<15.4f}{mejor:<12}{coincidencias_totales[i]}/5")

print("="*70)
print(f"{'PROMEDIO GLOBAL':<25}{np.mean(promedios_vgg16):<15.4f}{np.mean(promedios_vgg19):<15.4f}")
print(f"{'COINCIDENCIAS TOTALES':<25}{'':<15}{'':<15}{'':<12}{sum(coincidencias_totales)}/15")
print("="*70)

# Conclusion
ganador_general = "VGG16" if np.mean(promedios_vgg16) > np.mean(promedios_vgg19) else "VGG19"
print(f"\nCONCLUSION PREGUNTA 1:")
print(f"  - Modelo con mejores scores promedio: {ganador_general}")
print(f"  - Porcentaje de coincidencia global: {sum(coincidencias_totales)/15*100:.1f}%")
print(f"  - Ambos modelos producen recomendaciones visualmente coherentes")
print("="*70)

======================================================================
RESUMEN FINAL - PREGUNTA 1: BUSQUEDA DE IMAGENES SIMILARES
======================================================================

======================================================================
Ejemplo                  VGG16 Prom.    VGG19 Prom.    Mejor       Coincid.
======================================================================
Ejemplo 1 (idx=952)      0.6543         0.7036         VGG19       1/5
Ejemplo 2 (idx=396)      0.7521         0.7120         VGG16       2/5
Ejemplo 3 (idx=26)       0.7891         0.7833         VGG16       5/5
======================================================================
PROMEDIO GLOBAL          0.7318         0.7330         
COINCIDENCIAS TOTALES                                              8/15
======================================================================

CONCLUSION PREGUNTA 1:
  - Modelo con mejores scores promedio: VGG19
  - Porcentaje de coincidencia global: 53.3%
  - Ambos modelos producen recomendaciones visualmente coherentes
======================================================================

# ============================================================================
# ANALISIS DE PARAMETROS: VGG16 vs VGG19
# ============================================================================

print("="*70)
print("PREGUNTA 2: PARAMETROS ENTRENABLES")
print("="*70)

# --- Parametros VGG16 ---
print("\n" + "-"*70)
print("MODELO VGG16 (sin capa de clasificacion)")
print("-"*70)

total_params_vgg16 = feat_extractor_vgg16.count_params()
trainable_params_vgg16 = sum([np.prod(v.shape) for v in feat_extractor_vgg16.trainable_weights])
non_trainable_params_vgg16 = sum([np.prod(v.shape) for v in feat_extractor_vgg16.non_trainable_weights])

print(f"Total de parametros:        {total_params_vgg16:>15,}")
print(f"Parametros entrenables:     {trainable_params_vgg16:>15,}")
print(f"Parametros no entrenables:  {non_trainable_params_vgg16:>15,}")

# --- Parametros VGG19 ---
print("\n" + "-"*70)
print("MODELO VGG19 (sin capa de clasificacion)")
print("-"*70)

total_params_vgg19 = feat_extractor.count_params()
trainable_params_vgg19 = sum([np.prod(v.shape) for v in feat_extractor.trainable_weights])
non_trainable_params_vgg19 = sum([np.prod(v.shape) for v in feat_extractor.non_trainable_weights])

print(f"Total de parametros:        {total_params_vgg19:>15,}")
print(f"Parametros entrenables:     {trainable_params_vgg19:>15,}")
print(f"Parametros no entrenables:  {non_trainable_params_vgg19:>15,}")

# --- Diferencia ---
print("\n" + "="*70)
print("DIFERENCIA")
print("="*70)

diferencia_params = trainable_params_vgg19 - trainable_params_vgg16
porcentaje_aumento = (diferencia_params / trainable_params_vgg16) * 100

print(f"Diferencia absoluta:        {diferencia_params:>15,} parametros")
print(f"Incremento porcentual:      {porcentaje_aumento:>15.2f}%")
print("="*70)

======================================================================
PREGUNTA 2: PARAMETROS ENTRENABLES
======================================================================

----------------------------------------------------------------------
MODELO VGG16 (sin capa de clasificacion)
----------------------------------------------------------------------
Total de parametros:            134,260,544
Parametros entrenables:         134,260,544
Parametros no entrenables:                0

----------------------------------------------------------------------
MODELO VGG19 (sin capa de clasificacion)
----------------------------------------------------------------------
Total de parametros:            139,570,240
Parametros entrenables:         139,570,240
Parametros no entrenables:                0

======================================================================
DIFERENCIA
======================================================================
Diferencia absoluta:              5,309,696 parametros
Incremento porcentual:                 3.95%
======================================================================

# ============================================================================
# VISUALIZACION COMPARATIVA DE PARAMETROS
# ============================================================================

fig, axes = plt.subplots(1, 3, figsize=(18, 6), facecolor='#f8f9fa')

# --- Plot 1: Barras de parametros totales ---
ax = axes[0]
modelos = ['VGG16', 'VGG19']
params = [trainable_params_vgg16, trainable_params_vgg19]
colores = ['#4CAF50', '#2196F3']

bars = ax.bar(modelos, params, color=colores, edgecolor='black', linewidth=2)

ax.set_ylabel('Numero de Parametros', fontsize=12, fontweight='bold')
ax.set_title('Parametros Entrenables por Modelo', fontsize=14, fontweight='bold')
ax.yaxis.set_major_formatter(plt.FuncFormatter(lambda x, p: f'{int(x/1e6)}M'))
ax.grid(True, axis='y', alpha=0.3, linestyle='--')

for bar, param in zip(bars, params):
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 1e6,
            f'{param:,}', ha='center', va='bottom', fontsize=11, fontweight='bold')

# Añadir flecha indicando diferencia
ax.annotate('', xy=(1, trainable_params_vgg19), xytext=(1, trainable_params_vgg16),
            arrowprops=dict(arrowstyle='<->', color='red', lw=2))
ax.text(1.15, (trainable_params_vgg16 + trainable_params_vgg19)/2,
        f'+{diferencia_params:,}\n(+{porcentaje_aumento:.1f}%)',
        fontsize=10, fontweight='bold', color='red', va='center')

# --- Plot 2: Grafico de dona comparativo ---
ax = axes[1]

# Datos para dona
sizes = [trainable_params_vgg16, diferencia_params]
labels_dona = ['Parametros\ncomunes', 'Parametros\nadicionales\nVGG19']
colors_dona = ['#4CAF50', '#e74c3c']
explode = (0, 0.1)

wedges, texts, autotexts = ax.pie(sizes, explode=explode, labels=labels_dona, colors=colors_dona,
                                   autopct='%1.1f%%', startangle=90,
                                   textprops={'fontsize': 10, 'fontweight': 'bold'})
ax.set_title('Proporcion de Parametros Adicionales en VGG19', fontsize=14, fontweight='bold')

# --- Plot 3: Desglose por tipo de capa ---
ax = axes[2]

# Capas convolucionales vs fully connected (aproximado)
# VGG16: 13 conv + 2 FC (fc1, fc2) | VGG19: 16 conv + 2 FC
conv_vgg16 = 14714688  # Aproximado: capas convolucionales
fc_vgg16 = trainable_params_vgg16 - conv_vgg16

conv_vgg19 = 20024384  # Aproximado: capas convolucionales
fc_vgg19 = trainable_params_vgg19 - conv_vgg19

x = np.arange(2)
width = 0.35

bars1 = ax.bar(x - width/2, [conv_vgg16/1e6, conv_vgg19/1e6], width,
               label='Capas Convolucionales', color='#3498db', edgecolor='black')
bars2 = ax.bar(x + width/2, [fc_vgg16/1e6, fc_vgg19/1e6], width,
               label='Capas Fully Connected', color='#e67e22', edgecolor='black')

ax.set_ylabel('Parametros (Millones)', fontsize=12, fontweight='bold')
ax.set_title('Desglose por Tipo de Capa', fontsize=14, fontweight='bold')
ax.set_xticks(x)
ax.set_xticklabels(['VGG16', 'VGG19'], fontsize=12, fontweight='bold')
ax.legend(fontsize=10)
ax.grid(True, axis='y', alpha=0.3, linestyle='--')

plt.tight_layout()
plt.show()

# ============================================================================
# COMPARACION VISUAL DE ARQUITECTURAS
# ============================================================================

import matplotlib.patches as mpatches

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 10), facecolor='#f8f9fa')

# --- VGG16 Architecture ---
vgg16_blocks = [
    ('Input\n224x224x3', 0, '#ecf0f1', 0.3),
    ('Block 1\n2 Conv (64)', 2, '#E8F5E9', 0.6),
    ('Block 2\n2 Conv (128)', 2, '#C8E6C9', 0.6),
    ('Block 3\n3 Conv (256)', 3, '#A5D6A7', 0.9),
    ('Block 4\n3 Conv (512)', 3, '#81C784', 0.9),
    ('Block 5\n3 Conv (512)', 3, '#66BB6A', 0.9),
    ('Flatten', 0, '#ecf0f1', 0.3),
    ('FC1 (4096)', 1, '#4CAF50', 0.5),
    ('FC2 (4096)', 1, '#4CAF50', 0.5),
]

y_pos = 0
for block_name, num_layers, color, height in vgg16_blocks:
    rect = mpatches.FancyBboxPatch((0.5, y_pos), 2, height,
                                     boxstyle="round,pad=0.02,rounding_size=0.1",
                                     facecolor=color, edgecolor='black', linewidth=2)
    ax1.add_patch(rect)
    ax1.text(1.5, y_pos + height/2, block_name,
             ha='center', va='center', fontsize=10, fontweight='bold')
    y_pos += height + 0.15

ax1.set_xlim(0, 3)
ax1.set_ylim(-0.2, y_pos + 0.5)
ax1.axis('off')
ax1.set_title('VGG16\n13 Capas Conv + 2 FC\n' + f'{trainable_params_vgg16:,} parametros',
              fontsize=16, fontweight='bold', color='#4CAF50', pad=20)

# --- VGG19 Architecture ---
vgg19_blocks = [
    ('Input\n224x224x3', 0, '#ecf0f1', 0.3),
    ('Block 1\n2 Conv (64)', 2, '#E3F2FD', 0.6),
    ('Block 2\n2 Conv (128)', 2, '#BBDEFB', 0.6),
    ('Block 3\n4 Conv (256)', 4, '#90CAF9', 1.2),
    ('Block 4\n4 Conv (512)', 4, '#64B5F6', 1.2),
    ('Block 5\n4 Conv (512)', 4, '#42A5F5', 1.2),
    ('Flatten', 0, '#ecf0f1', 0.3),
    ('FC1 (4096)', 1, '#2196F3', 0.5),
    ('FC2 (4096)', 1, '#2196F3', 0.5),
]

y_pos = 0
for block_name, num_layers, color, height in vgg19_blocks:
    rect = mpatches.FancyBboxPatch((0.5, y_pos), 2, height,
                                     boxstyle="round,pad=0.02,rounding_size=0.1",
                                     facecolor=color, edgecolor='black', linewidth=2)
    ax2.add_patch(rect)
    ax2.text(1.5, y_pos + height/2, block_name,
             ha='center', va='center', fontsize=10, fontweight='bold')
    y_pos += height + 0.15

ax2.set_xlim(0, 3)
ax2.set_ylim(-0.2, y_pos + 0.5)
ax2.axis('off')
ax2.set_title('VGG19\n16 Capas Conv + 2 FC\n' + f'{trainable_params_vgg19:,} parametros',
              fontsize=16, fontweight='bold', color='#2196F3', pad=20)

# Titulo general
fig.suptitle('COMPARACION DE ARQUITECTURAS', fontsize=20, fontweight='bold', y=0.98)

plt.tight_layout()
plt.show()

# Diferencias clave
print("\n" + "="*70)
print("DIFERENCIAS ARQUITECTONICAS CLAVE")
print("="*70)
print("""
VGG16:                              VGG19:
- Block 1: 2 capas conv (64)        - Block 1: 2 capas conv (64)
- Block 2: 2 capas conv (128)       - Block 2: 2 capas conv (128)
- Block 3: 3 capas conv (256)       - Block 3: 4 capas conv (256)  [+1]
- Block 4: 3 capas conv (512)       - Block 4: 4 capas conv (512)  [+1]
- Block 5: 3 capas conv (512)       - Block 5: 4 capas conv (512)  [+1]
- FC1: 4096 neuronas                - FC1: 4096 neuronas
- FC2: 4096 neuronas                - FC2: 4096 neuronas

Total: 13 conv + 2 FC               Total: 16 conv + 2 FC
      = 15 capas                          = 18 capas
""")
print("="*70)

======================================================================
DIFERENCIAS ARQUITECTONICAS CLAVE
======================================================================

VGG16:                              VGG19:
- Block 1: 2 capas conv (64)        - Block 1: 2 capas conv (64)
- Block 2: 2 capas conv (128)       - Block 2: 2 capas conv (128)
- Block 3: 3 capas conv (256)       - Block 3: 4 capas conv (256)  [+1]
- Block 4: 3 capas conv (512)       - Block 4: 4 capas conv (512)  [+1]
- Block 5: 3 capas conv (512)       - Block 5: 4 capas conv (512)  [+1]
- FC1: 4096 neuronas                - FC1: 4096 neuronas
- FC2: 4096 neuronas                - FC2: 4096 neuronas

Total: 13 conv + 2 FC               Total: 16 conv + 2 FC
      = 15 capas                          = 18 capas

======================================================================

# ============================================================================
# RESPUESTA FINAL - PREGUNTA 2
# ============================================================================

print("\n")
print("+"*70)
print("|{:^68}|".format("RESPUESTA PREGUNTA 2"))
print("+"*70)

print(f"""
|{'':^68}|
|{'CANTIDAD DE PARAMETROS ENTRENABLES':^68}|
|{'(sin ultima capa de clasificacion)':^68}|
|{'':^68}|
|{'-'*68}|
|{'Arquitectura':<20}|{'Parametros Entrenables':>46}|
|{'-'*68}|
|{'VGG16':<20}|{trainable_params_vgg16:>40,}      |
|{'VGG19':<20}|{trainable_params_vgg19:>40,}      |
|{'-'*68}|
|{'DIFERENCIA':<20}|{diferencia_params:>40,}      |
|{'INCREMENTO':<20}|{porcentaje_aumento:>39.2f}%      |
|{'-'*68}|
""")

print("+"*70)
print("|{:^68}|".format("CONCLUSION"))
print("+"*70)
print(f"|{'':^68}|")
print(f"|  {'VGG19 tiene MAS parametros entrenables que VGG16':<64}  |")
print(f"|{'':^68}|")
print(f"|  {'Justificacion:':<64}  |")
print(f"|  {'VGG19 posee 3 capas convolucionales adicionales:':<64}  |")
print(f"|    {'- 1 capa extra en Block 3 (256 filtros)':<62}  |")
print(f"|    {'- 1 capa extra en Block 4 (512 filtros)':<62}  |")
print(f"|    {'- 1 capa extra en Block 5 (512 filtros)':<62}  |")
print(f"|{'':^68}|")
print(f"|  {'Estas capas adicionales agregan ~5.1% mas parametros.':<64}  |")
print(f"|{'':^68}|")
print("+"*70)


++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
|                        RESPUESTA PREGUNTA 2                        |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

|                                                                    |
|                 CANTIDAD DE PARAMETROS ENTRENABLES                 |
|                 (sin ultima capa de clasificacion)                 |
|                                                                    |
|--------------------------------------------------------------------|
|Arquitectura        |                        Parametros Entrenables|
|--------------------------------------------------------------------|
|VGG16               |                             134,260,544      |
|VGG19               |                             139,570,240      |
|--------------------------------------------------------------------|
|DIFERENCIA          |                               5,309,696      |
|INCREMENTO          |                                   3.95%      |
|--------------------------------------------------------------------|

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
|                             CONCLUSION                             |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
|                                                                    |
|  VGG19 tiene MAS parametros entrenables que VGG16                  |
|                                                                    |
|  Justificacion:                                                    |
|  VGG19 posee 3 capas convolucionales adicionales:                  |
|    - 1 capa extra en Block 3 (256 filtros)                         |
|    - 1 capa extra en Block 4 (512 filtros)                         |
|    - 1 capa extra en Block 5 (512 filtros)                         |
|                                                                    |
|  Estas capas adicionales agregan ~5.1% mas parametros.             |
|                                                                    |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

# ============================================================================
# VISUALIZACION RESUMEN FINAL
# ============================================================================

fig, ax = plt.subplots(figsize=(12, 8), facecolor='white')
ax.axis('off')

# Titulo
ax.text(0.5, 0.95, 'RESPUESTA PREGUNTA 2', fontsize=24, fontweight='bold',
        ha='center', va='top', color='#2c3e50')
ax.text(0.5, 0.88, 'Comparacion de Parametros Entrenables', fontsize=16,
        ha='center', va='top', color='#7f8c8d', style='italic')

# Cajas de modelos
# VGG16
rect_vgg16 = mpatches.FancyBboxPatch((0.08, 0.50), 0.35, 0.30,
                                       boxstyle="round,pad=0.02,rounding_size=0.05",
                                       facecolor='#e8f5e9', edgecolor='#4CAF50', linewidth=4)
ax.add_patch(rect_vgg16)
ax.text(0.255, 0.73, 'VGG16', fontsize=20, fontweight='bold', ha='center', color='#4CAF50')
ax.text(0.255, 0.63, f'{trainable_params_vgg16:,}', fontsize=16, fontweight='bold', ha='center', color='#2c3e50')
ax.text(0.255, 0.55, 'parametros', fontsize=12, ha='center', color='#7f8c8d')

# VGG19
rect_vgg19 = mpatches.FancyBboxPatch((0.57, 0.50), 0.35, 0.30,
                                       boxstyle="round,pad=0.02,rounding_size=0.05",
                                       facecolor='#e3f2fd', edgecolor='#2196F3', linewidth=4)
ax.add_patch(rect_vgg19)
ax.text(0.745, 0.73, 'VGG19', fontsize=20, fontweight='bold', ha='center', color='#2196F3')
ax.text(0.745, 0.63, f'{trainable_params_vgg19:,}', fontsize=16, fontweight='bold', ha='center', color='#2c3e50')
ax.text(0.745, 0.55, 'parametros', fontsize=12, ha='center', color='#7f8c8d')

# Flecha y diferencia
ax.annotate('', xy=(0.55, 0.65), xytext=(0.45, 0.65),
            arrowprops=dict(arrowstyle='->', color='#e74c3c', lw=3))
ax.text(0.50, 0.72, f'+{diferencia_params:,}', fontsize=14, fontweight='bold',
        ha='center', color='#e74c3c')
ax.text(0.50, 0.68, f'(+{porcentaje_aumento:.1f}%)', fontsize=11, ha='center', color='#e74c3c')

# Caja de conclusion
rect_conclusion = mpatches.FancyBboxPatch((0.10, 0.08), 0.80, 0.35,
                                            boxstyle="round,pad=0.02,rounding_size=0.05",
                                            facecolor='#fff9c4', edgecolor='#f9a825', linewidth=3)
ax.add_patch(rect_conclusion)

ax.text(0.50, 0.38, 'CONCLUSION', fontsize=16, fontweight='bold', ha='center', color='#f57f17')
ax.text(0.50, 0.32, 'VGG19 tiene MAS parametros entrenables', fontsize=14, fontweight='bold',
        ha='center', color='#2c3e50')
ax.text(0.50, 0.24, 'La diferencia se debe a 3 capas convolucionales adicionales:', fontsize=11,
        ha='center', color='#5d4037')
ax.text(0.50, 0.18, '• +1 capa en Block 3 (256 filtros)    • +1 capa en Block 4 (512 filtros)    • +1 capa en Block 5 (512 filtros)',
        fontsize=10, ha='center', color='#5d4037')
ax.text(0.50, 0.12, f'Esto representa un incremento del {porcentaje_aumento:.1f}% en parametros totales.',
        fontsize=11, ha='center', color='#5d4037', style='italic')

ax.set_xlim(0, 1)
ax.set_ylim(0, 1)

plt.tight_layout()
plt.show()

Aspecto	Descripcion
Tema	Recomendacion basada en similitud visual usando CNNs
Arquitecturas	VGG16 y VGG19
Dataset	Fashion Images
Formato de entrega	Notebook
Fecha de entrega	30 de Noviembre 2025

	images/4_9_037.png	images/6_7_002.png	images/0_0_065.png	images/2_2_016.png	images/3_0_051.png	images/4_6_073.png	images/6_4_014.png	images/1_4_023.png	images/4_1_031.png	images/0_2_008.png	...	images/4_9_001.png	images/1_7_016.png	images/1_5_030.png	images/0_0_050.png	images/6_2_051.png	images/5_0_041.png	images/2_1_024.png	images/6_2_036.png	images/1_0_045.png	images/1_6_028.png
images/4_9_037.png	1.000000	0.234604	0.270189	0.249941	0.255349	0.680604	0.349751	0.518958	0.241923	0.277481	...	0.351203	0.430215	0.195535	0.333183	0.391848	0.224119	0.262152	0.234611	0.259536	0.518413
images/6_7_002.png	0.234604	1.000000	0.511521	0.383287	0.462207	0.206403	0.289353	0.319520	0.146800	0.325533	...	0.213976	0.380928	0.247740	0.395100	0.259345	0.623821	0.177394	0.275122	0.477618	0.438690
images/0_0_065.png	0.270189	0.511521	1.000000	0.334260	0.428377	0.245892	0.404932	0.334245	0.148620	0.290298	...	0.207807	0.354473	0.149409	0.627904	0.215543	0.551125	0.156673	0.254642	0.574104	0.470782
images/2_2_016.png	0.249941	0.383287	0.334260	1.000000	0.282075	0.236036	0.381667	0.353902	0.113987	0.393667	...	0.294710	0.285929	0.297188	0.299190	0.449039	0.308903	0.175529	0.516861	0.260724	0.333543
images/3_0_051.png	0.255349	0.462207	0.428377	0.282075	1.000000	0.230695	0.282458	0.370668	0.177571	0.308622	...	0.256329	0.337162	0.219169	0.492444	0.306134	0.366985	0.283139	0.305349	0.574615	0.477951
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
images/5_0_041.png	0.224119	0.623821	0.551125	0.308903	0.366985	0.223562	0.258447	0.245701	0.145548	0.340220	...	0.174923	0.386429	0.190107	0.397802	0.247668	1.000000	0.237673	0.234311	0.402411	0.414470
images/2_1_024.png	0.262152	0.177394	0.156673	0.175529	0.283139	0.217404	0.197698	0.252936	0.313078	0.255856	...	0.388821	0.236676	0.313779	0.170812	0.303209	0.237673	1.000000	0.160127	0.186704	0.348633
images/6_2_036.png	0.234611	0.275122	0.254642	0.516861	0.305349	0.256257	0.418394	0.341404	0.130103	0.325769	...	0.411298	0.284604	0.178414	0.379742	0.590484	0.234311	0.160127	1.000000	0.235376	0.232773
images/1_0_045.png	0.259536	0.477618	0.574104	0.260724	0.574615	0.271201	0.390164	0.399710	0.178353	0.262325	...	0.220495	0.390571	0.158452	0.443267	0.238740	0.402411	0.186704	0.235376	1.000000	0.440751
images/1_6_028.png	0.518413	0.438690	0.470782	0.333543	0.477951	0.340128	0.419501	0.471212	0.166863	0.323689	...	0.436489	0.613510	0.273013	0.434295	0.357210	0.414470	0.348633	0.232773	0.440751	1.000000

	images/4_9_037.png	images/6_7_002.png	images/0_0_065.png	images/2_2_016.png	images/3_0_051.png	images/4_6_073.png	images/6_4_014.png	images/1_4_023.png	images/4_1_031.png	images/0_2_008.png	...	images/4_9_001.png	images/1_7_016.png	images/1_5_030.png	images/0_0_050.png	images/6_2_051.png	images/5_0_041.png	images/2_1_024.png	images/6_2_036.png	images/1_0_045.png	images/1_6_028.png
images/4_9_037.png	1.000000	0.246062	0.283663	0.266740	0.241140	0.594259	0.313979	0.483542	0.239732	0.318703	...	0.396407	0.393539	0.213051	0.245248	0.428944	0.282033	0.237320	0.278771	0.318151	0.486188
images/6_7_002.png	0.246062	1.000000	0.550363	0.347476	0.451177	0.216592	0.325017	0.351115	0.173656	0.358070	...	0.284720	0.429984	0.215832	0.346894	0.304129	0.532368	0.097576	0.306875	0.576793	0.355692
images/0_0_065.png	0.283663	0.550363	1.000000	0.410117	0.451497	0.263972	0.418978	0.353457	0.150228	0.429977	...	0.260815	0.400474	0.195367	0.653091	0.311516	0.648227	0.109908	0.365587	0.605920	0.426587
images/2_2_016.png	0.266740	0.347476	0.410117	1.000000	0.316359	0.275954	0.434996	0.361416	0.134597	0.459089	...	0.321647	0.380279	0.265330	0.311073	0.471006	0.331293	0.130500	0.546710	0.356725	0.317658
images/3_0_051.png	0.241140	0.451177	0.451497	0.316359	1.000000	0.197331	0.286894	0.356369	0.191200	0.362173	...	0.276441	0.350337	0.176082	0.462562	0.366559	0.322122	0.170922	0.349192	0.551198	0.399418
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
images/5_0_041.png	0.282033	0.532368	0.648227	0.331293	0.322122	0.240706	0.356089	0.300686	0.142373	0.408893	...	0.207521	0.381667	0.196213	0.459438	0.303401	1.000000	0.115555	0.286341	0.487723	0.315371
images/2_1_024.png	0.237320	0.097576	0.109908	0.130500	0.170922	0.164226	0.162832	0.206901	0.331468	0.188610	...	0.425571	0.177083	0.226012	0.129678	0.234193	0.115555	1.000000	0.154881	0.144900	0.201714
images/6_2_036.png	0.278771	0.306875	0.365587	0.546710	0.349192	0.283427	0.584111	0.398095	0.149442	0.389695	...	0.371073	0.321338	0.282947	0.405400	0.653835	0.286341	0.154881	1.000000	0.328660	0.238560
images/1_0_045.png	0.318151	0.576793	0.605920	0.356725	0.551198	0.295753	0.457204	0.468110	0.132301	0.363992	...	0.366497	0.552081	0.196607	0.514968	0.337822	0.487723	0.144900	0.328660	1.000000	0.487877
images/1_6_028.png	0.486188	0.355692	0.426587	0.317658	0.399418	0.395290	0.358418	0.441232	0.154469	0.349450	...	0.416565	0.513014	0.237698	0.404568	0.310631	0.315371	0.201714	0.238560	0.487877	1.000000

Arquitectura	Parámetros Entrenables
VGG16	119,545,856
VGG19	125,641,216
Diferencia	6,095,360 (+5.1%)

Magister en Inteligencia Artificial¶

Sistemas de Recomendacion¶

Equipo Docente¶

Practico Clase 6: Deep Learning para Recomendacion de Imagenes¶

Detalles de la Actividad¶

Alumno¶

Javier Augusto Rebull Saucedo¶

Objetivo del Practico¶

📖 Introducción¶

Práctico Deep Learning para Recomendación¶

Importación de Librerías¶

descarga de imagenes¶

Ejercicio con vgg19¶

cargamos la CNN pre-entrenada en ImageNet¶

procesamiento de imágenes para dárselas como input a la CNN¶

galería de imagenes¶

Visualización del Preprocesamiento de Imágenes¶

Extraccion de Features con VGG19 (Imagen Individual)¶

Repetimos el Ejericio con vgg16¶

cargamos la CNN pre-entrenada en ImageNet¶

Extraccion de Features con VGG16 (Imagen Individual)¶

ACTIVIDAD¶

Preguntas a Resolver¶

¶

Pregunta 1: Busqueda de Imagenes Similares con VGG16 y VGG19 (3 ptos)¶

Observaciones Pregunta 1¶

Pregunta 2: Comparacion de Parametros Entrenables VGG16 vs VGG19 (2 ptos)¶

Respuesta Pregunta 2¶

Pregunta 2: Conclusiones y Reflexión Personal¶

Conclusiones Principales¶

Efectividad del Transfer Learning¶

Comparación Técnica: VGG16 vs. VGG19¶

Aplicabilidad en la Industria¶

Reflexión Personal¶

Referencias¶

🎓 Magister en Inteligencia Artificial¶

Pontificia Universidad Católica de Chile¶